A computer external device quality evaluation method and system

Abstract

The application discloses a kind of computer external device quality evaluation method and system, it is related to external device quality evaluation technical field, including the following steps, the time series data of thermal printer is obtained, interpolation is handled to the missing value of time series data using adaptive mixed interpolation strategy, obtain time series data matrix, to time series data matrix, carry out adaptive phased data standardization processing, obtain standardization characteristic sequence, to standardization characteristic sequence, carry out health factor dynamic selection processing, obtain health factor dataset, to health factor dataset carries out adaptive data division, obtain the dataset after division, based on first dataset and second dataset, using improved AGA-GRU-Informer model assesses and predicts the health state of thermal printer, realize the significant improvement of thermal printer health state assessment accuracy, adaptability and timeliness, and enhance the generalization ability of model between different devices.

Description

Technical Field

[0001] This invention relates to the field of peripheral device quality assessment technology, and in particular to a method and system for assessing the quality of computer peripheral devices. Background Technology

[0002] In recent years, with the widespread application of computer peripherals in key areas such as healthcare, retail, and logistics, thermal printers, as important data output devices, have their operational reliability and quality status directly affecting the continuity and stability of the entire system. Therefore, accurate and efficient health status assessment and prediction of thermal printers has become an important technical means to improve equipment management and achieve predictive maintenance.

[0003] However, the monitoring data generated by thermal printers in actual operation are characterized by strong time series, multi-source heterogeneity, and a large amount of random missing data and noise. Traditional assessment methods often fail to fully explore the dynamic evolution and health degradation characteristics, resulting in insufficient real-time performance and accuracy of condition assessment. Existing technologies usually adopt fixed interpolation methods and data standardization strategies, ignoring the differences in data distribution at different stages of the equipment's life cycle. Furthermore, feature selection is mostly static and global, which cannot adapt to the dynamic changes of key health factors during equipment aging. In addition, most prediction models have limitations in capturing the long-term dependence and local dynamics of time series data, and fail to effectively integrate short-term condition fluctuations and long-term degradation trends, resulting in lagging or significantly biased assessment results. Summary of the Invention

[0004] The technical problem solved by this invention is that: the monitoring data of thermal printers throughout their entire life cycle is incomplete, has a drifting distribution, and exhibits dynamic changes in key health factors. Existing fixed interpolation methods and static feature selection strategies cannot adapt to the data characteristics of different operating stages. Existing prediction models are unable to effectively integrate short-term operational fluctuations with long-term degradation trends, resulting in lagging and inaccurate health status assessment results.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for evaluating the quality of computer peripherals, comprising the following steps:

[0006] Step S1: Obtain the time series data of the thermal printer, use an adaptive hybrid interpolation strategy to interpolate the missing values ​​of the time series data to obtain a time series data matrix, and perform adaptive phased data standardization processing on the time series data matrix to obtain a standardized feature sequence.

[0007] Step S2: Perform dynamic selection of health factors on the standardized feature sequence to obtain a health factor dataset. Perform adaptive data partitioning on the health factor dataset to obtain the partitioned dataset.

[0008] Step S3: Based on the first and second datasets, the improved AGA-GRU-Informer model is used to assess and predict the health status of the thermal printer.

[0009] As a preferred embodiment of the computer peripheral device quality assessment method of the present invention, step S1 includes steps S101, S102 and S103.

[0010] Step S101: Obtain the time-series data collected in chronological order during the operation of the thermal printer, as well as the corresponding maintenance and fault record data;

[0011] Time-series data includes monitoring data and attribute data;

[0012] The monitoring data includes printhead temperature data, motor current data, printing speed data, and paper transport status data.

[0013] The attribute data includes the thermal printer's runtime data and the thermal printer's ID;

[0014] Specifically, the paper transfer status data includes 0 for normal and 1 for paper jam.

[0015] Step S102: For time series data, the timestamp of the collected time series data is combined with the ID of the thermal printer as a unique identifier to identify and delete duplicate records, thereby obtaining deduplicated time series data.

[0016] If there are missing values ​​in the deduplicated time series data, an adaptive hybrid interpolation strategy is used to interpolate the missing values ​​to obtain the time series data matrix.

[0017] The adaptive hybrid interpolation strategy specifically includes selecting 10 samples before and after each data point with missing values ​​in the deduplicated time series data to form a local data sequence, calculating the variance of the local data sequence, and obtaining the local data sequence variance.

[0018] The variance of the local data sequence is compared with a preset missing threshold. If the variance of the local data sequence is less than the missing threshold, linear interpolation is used to interpolate the missing values.

[0019] If the variance of a local data sequence is greater than or equal to the missing threshold, then quadratic polynomial interpolation is used to interpolate the missing values.

[0020] After deduplication and interpolation, the time series data are arranged in the order of the timestamps of the collected time series data to obtain a time series data matrix.

[0021] Time series data matrix representation is ;

[0022] in, Indicates the number of samples. Indicates the quantity of monitoring data. The timestamp, thermal printer ID, and runtime data represent the time series data after deduplication and interpolation.

[0023] As a preferred embodiment of the computer peripheral device quality assessment method of the present invention, step S103, adaptive phased data standardization processing specifically includes dividing the thermal printer's running time data in the time series data matrix according to interval division logic to obtain the corresponding life cycle stage;

[0024] The life cycle stages include the first stage, the second stage, and the third stage;

[0025] The interval division logic specifically includes dividing the data into the first stage when the running time is less than the first running threshold.

[0026] When the runtime data is greater than or equal to the first runtime threshold and less than the second runtime threshold, it is classified into the second stage;

[0027] When the runtime data is greater than or equal to the second runtime threshold, it is divided into the third stage;

[0028] Calculate the mean and standard deviation of all monitoring data in the first, second and third stages respectively;

[0029] For each data point in the time series data matrix, the corresponding life cycle stage is determined based on the corresponding running time data, and the standardization calculation is performed based on the mean and standard deviation corresponding to the life cycle stage. The calculation logic is as follows: the difference between the monitoring data of the data point and the corresponding mean is calculated to obtain the deviation value, and the ratio of the deviation value to the corresponding standard deviation is calculated to obtain the standardized monitoring data.

[0030] A standard range is set, and monitoring data exceeding the standard range is replaced with the boundary value corresponding to the standard range.

[0031] The processed standardized monitoring data, the timestamps of the deduplicated and interpolated time series data, the thermal printer ID, and the running time data are combined to form a standardized feature matrix;

[0032] The standardized feature matrix is ​​arranged in the order of the timestamps of the time series data after deduplication and interpolation to form a standardized feature sequence.

[0033] In a preferred embodiment of the computer peripheral device quality assessment method of the present invention, step S2 includes steps S201 and S202.

[0034] Step S201: Extract maintenance events and their corresponding timestamps from the maintenance records of the thermal printer and fault events and their corresponding timestamps from the fault history data. Align the timestamps corresponding to the maintenance events and fault events with the timestamps of the time-series data after deduplication and interpolation in the standardized feature sequence.

[0035] For each data point in the standardized feature sequence, the health status value of the thermal printer corresponding to each data point in the standardized feature sequence is calculated by linear interpolation between the maintenance events and fault events adjacent to the data point. This health status value is used as the actual health value. The range of the health status value is 0 to 1, where 1 indicates that the thermal printer is completely healthy and 0 indicates that the thermal printer is completely unresponsive.

[0036] The calculated true health status value of each data point is arranged according to the timestamp order of the time series data after deduplication and interpolation in the corresponding standardized feature sequence to form a true health status value sequence.

[0037] Based on interval partitioning logic, the corresponding life cycle stages are obtained. The monitoring data in the standardized feature sequence are grouped according to the life cycle stages, and the correlation coefficient between each monitoring data and the real health status value sequence in each life cycle stage is calculated. Specifically, the Pearson correlation coefficient between the printhead temperature data, motor current data, printing speed data, and running time data and the real health status value is calculated, and the point-to-binary correlation coefficient between the paper transfer status data and the real health status value is calculated.

[0038] As a preferred embodiment of the computer peripheral device quality assessment method described in this invention, for the first, second and third stages of the life cycle, data that meets the conditions are selected from the monitoring data as health factors for the corresponding life cycle stages according to preset screening conditions.

[0039] For the first stage, the specific screening criteria include that the absolute value of the correlation coefficient is greater than the first correlation threshold, and the monitored data belongs to printing speed data, printhead temperature data, or motor current data.

[0040] For the second stage, the specific screening criteria include that the absolute value of the correlation coefficient is greater than the first correlation threshold, and the monitoring data belongs to the temperature data of the print head, the motor current data, or the paper transfer status data.

[0041] For the third stage, the specific screening criteria include that the absolute value of the correlation coefficient is greater than the first correlation threshold, and that the monitoring data belongs to motor current data, running time data, or printhead temperature data.

[0042] Based on the absolute value of the correlation coefficient of each health factor in the corresponding life cycle stage, a weight is assigned to each health factor. The initial weight of the health factor is equal to the absolute value of the corresponding correlation coefficient. The weights of all health factors in the same life cycle stage are normalized to obtain and output the first stage health factor set and corresponding weights, the second stage health factor set and corresponding weights, and the third stage health factor set and corresponding weights.

[0043] The first-stage health factor set, the second-stage health factor set, and the third-stage health factor set constitute the health factor dataset.

[0044] As a preferred embodiment of the computer peripheral device quality assessment method of the present invention, step S202 involves adaptively partitioning the health factor dataset to obtain the partitioned dataset, specifically the first dataset and the second dataset.

[0045] The samples in the health factor dataset are sorted in ascending order by the timestamps of the time series data after deduplication and interpolation, and then sorted according to the corresponding life cycle stage to obtain the first stage data sequence, the second stage data sequence, and the third stage data sequence.

[0046] Adaptive data partitioning specifically includes using a single decision strategy to divide all thermal printers into independent and non-overlapping datasets;

[0047] Based on the data from the first, second, and third phases, a corresponding sliding window is set for each lifecycle phase, including the new machine window, the stable period window, and the aging period window.

[0048] The first phase corresponds to the new opportunity window, with a window length of 100 samples;

[0049] The second phase corresponds to the stable period window, with a window length of 300 samples;

[0050] The third stage corresponds to the aging period window, with a window length of 500 samples;

[0051] A corresponding sliding window is generated for the first stage data sequence, the second stage data sequence, and the third stage data sequence. When generating the sliding window, the overlap ratio between adjacent windows is set as the window overlap threshold.

[0052] All thermal printers are divided into a first dataset and a second dataset according to a preset ratio;

[0053] The first dataset includes a health factor dataset of a first proportion, and the second dataset includes a health factor dataset of a second proportion.

[0054] As a preferred embodiment of the computer peripheral device quality assessment method of the present invention, step S3 includes steps S301, S302, S303, S304 and S305.

[0055] Step S301: An adaptive genetic algorithm is used to automatically optimize the hidden layer dimension parameters of the GRU layer to obtain the optimal hidden layer dimension.

[0056] Step S301 includes steps S3011, S3012, S3013, S3014, S3015, S3016 and S3017;

[0057] Step S3011: Within the preset range of hidden layer dimensions, a set of candidate dimension values ​​is randomly generated to form the initial population;

[0058] Repeat steps S3012 to S3016 until the preset maximum number of iterations is reached to obtain the final generation of candidate dimension value population;

[0059] Step S3012: For each candidate dimension value in the current population, repeat steps S3013 to S3015 to obtain the corresponding fitness.

[0060] Step S3013: Construct a simplified GRU network, where the hidden layer dimensions of the simplified GRU network are candidate dimension values;

[0061] Step S3014: Sort all samples in the first dataset in ascending order according to the timestamps of the time series data after deduplication and interpolation, and select the third proportion of samples with the earliest timestamps in the first dataset to form the evaluation data sequence.

[0062] The evaluation data sequence is divided into continuous time-series samples using a fixed-length sliding window;

[0063] Step S3015: Input the time series samples into a simplified GRU network for processing to obtain the predicted health status sequence;

[0064] Calculate the root mean square of the difference between the predicted health status sequence and the corresponding true health status value sequence, and use it as the root mean square error of the candidate dimension value.

[0065] The reciprocal of the root mean square error is calculated as the fitness of the current candidate dimension value;

[0066] Step S3016: Based on the fitness of all candidate dimension values ​​in the current population, perform selection, crossover, and mutation operations to generate a new generation of candidate dimension value population;

[0067] Step S3017: From the final generation of candidate dimension values, output the candidate dimension value with the highest fitness as the optimal hidden layer dimension.

[0068] The selection process specifically includes sorting all candidate dimension values ​​according to their fitness and retaining the fourth-highest fitness candidate dimension values ​​as the parent individuals.

[0069] The crossover specifically involves randomly selecting two candidate dimension values ​​from the parent individuals, denoted as... and The offspring dimension values ​​are generated through linear combination, expressed as follows:

[0070] ;

[0071] in, Represents the child dimension value. Represents a random number in the range [0,1].

[0072] The mutation specifically involves adaptively perturbing the generated child dimension values ​​to obtain the final child dimension values, expressed as follows:

[0073] ;

[0074] in, This represents the final child dimension value. For adaptive coefficient of variation, This represents normally distributed noise. This indicates that the mean is 0 and the variance is 0. The normally distributed noise, after variation, will Constraints Within the range.

[0075] In a preferred embodiment of the computer peripheral device quality assessment method of the present invention, step S302 involves extracting feature sequence fragments from the first dataset and the second dataset according to the life cycle stage and the corresponding sliding window of the sample.

[0076] For each feature sequence segment, extract a continuous sequence segment with a preset step length following the feature sequence segment from the corresponding real health status value sequence, and use the continuous sequence segment as the target sequence of the feature sequence segment;

[0077] Based on the feature sequence fragments and the corresponding target sequences, the first time series sample and the second time series sample are generated respectively.

[0078] The process of generating the time-series samples includes adding sine-cosine position codes to the feature sequence segments and the corresponding target sequences, respectively.

[0079] The first and second time series samples are processed by a gated recurrent unit layer to obtain the corresponding GRU hidden state sequences.

[0080] Step S303: Normalize each GRU hidden state vector in the GRU hidden state sequence to obtain a normalized feature sequence;

[0081] The improved AGA-GRU-Informer model employs three independent linear transformation layers to map the normalized feature sequence into a query matrix, a key matrix, and a value matrix, respectively.

[0082] Based on the improved AGA-GRU-Informer model, the ProbSparse self-attention mechanism uses the query matrix and key matrix to filter queries and obtain a simplified query matrix. Specifically, for each query vector in the query matrix, the similarity distribution between the query vector and the set of all key vectors in the key matrix is ​​calculated.

[0083] Based on the similarity distribution, the sparsity score of each query vector is calculated;

[0084] Based on the sparsity score, the top [number] [items] are selected from the query matrix. A number of important query vectors are used to form a simplified query matrix, where u is the preset number of important queries;

[0085] Step S304, simplify the query matrix Key matrix Sum matrix The input is fed into the ProbSparse attention computation layer of the improved AGA-GRU-Informer model, and the long-term dependency feature representation is obtained by computation using the standard dot product attention formula.

[0086] As a preferred embodiment of the computer peripheral device quality assessment method described in this invention, in step S305, the long-term dependency feature representation and the sine-cosine position encoding of the target sequence are input into the generative decoder of the improved AGA-GRU-Informer model. The generative decoder processes the data through a three-layer self-attention mechanism and a fully connected layer, and outputs a context aggregation representation.

[0087] The context aggregation representation is processed by linear transformation and activation function to output a predicted sequence of the health status of the thermal printer.

[0088] Based on the mapping of context aggregation representation, the predicted health status value of the thermal printer is obtained. Specifically, this includes extracting the feature vector of the starting position of the target sequence in the context aggregation representation, and mapping the feature vector to an unscaled original prediction value through a fully connected layer.

[0089] The original predicted value is input into the Sigmoid activation function for processing to obtain the final health status prediction value.

[0090] The predicted value ranges from 0 to 1, where 1 indicates that the thermal printer is fully healthy and 0 indicates that it is completely unusable.

[0091] A computer peripheral device quality assessment system, comprising a processing module, a partitioning module, and an assessment module;

[0092] The processing module acquires the time-series data of the thermal printer, uses an adaptive hybrid interpolation strategy to interpolate the missing values ​​of the time-series data to obtain a time-series data matrix, and performs adaptive phased data standardization processing on the time-series data matrix to obtain a standardized feature sequence.

[0093] The partitioning module performs dynamic selection of health factors on the standardized feature sequences to obtain a health factor dataset, and then performs adaptive data partitioning on the health factor dataset to obtain the partitioned dataset.

[0094] The evaluation module, based on the first and second datasets, uses an improved AGA-GRU-Informer model to evaluate and predict the health status of thermal printers.

[0095] The beneficial effects of this invention are as follows: By employing an adaptive hybrid interpolation strategy and phased data standardization, it effectively addresses the dynamic changes in data distribution throughout the entire lifecycle of the equipment, significantly improving the quality and consistency of input data and laying a reliable foundation for subsequent analysis. Furthermore, the proposed dynamic health factor selection mechanism can automatically filter and assign weights to key monitoring parameters most relevant to the health status based on different operating stages of the equipment, solving the fundamental problem that traditional static feature sets cannot adapt to the equipment degradation process. The improved AGA-GRU-Informer hybrid model optimizes the network structure through an adaptive genetic algorithm and combines the ability of gated recurrent units to capture short-term dynamics with the ability of an Informer model integrating ProbSparse self-attention to capture long-term dependencies, achieving more accurate fitting and prediction of complex degradation trends. This solution forms a closed-loop optimization from data preprocessing and feature construction to the prediction model, ultimately significantly improving the accuracy, adaptability, and timeliness of thermal printer health status assessment and enhancing the model's generalization ability across different devices. Attached Figure Description

[0096] Figure 1 This is a flowchart illustrating the steps of a method for evaluating the quality of computer peripherals according to an embodiment of the present invention.

[0097] Figure 2This is a basic flowchart of a computer peripheral device quality assessment system provided in one embodiment of the present invention. Detailed Implementation

[0098] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0099] Example 1, referring to Figure 1 As an embodiment of the present invention, a method for evaluating the quality of computer peripheral devices is provided, comprising the following steps:

[0100] Step S1: Obtain the time series data of the thermal printer, use an adaptive hybrid interpolation strategy to interpolate the missing values ​​of the time series data to obtain a time series data matrix, and perform adaptive phased data standardization processing on the time series data matrix to obtain a standardized feature sequence.

[0101] Step S2: Perform dynamic selection of health factors on the standardized feature sequence to obtain a health factor dataset. Perform adaptive data partitioning on the health factor dataset to obtain the partitioned dataset.

[0102] Step S3: Based on the first and second datasets, the improved AGA-GRU-Informer model is used to assess and predict the health status of the thermal printer.

[0103] In one embodiment, this design framework targets thermal printers, a specific external device widely used in industrial automation, logistics sorting, and commercial transactions. Steps S1, S2, and S3 constitute a complete closed-loop processing flow. Such external devices require continuous operation 24 / 7; their failures directly lead to business process interruptions and significant losses. Therefore, high-precision predictive maintenance of their health status is urgently needed. Step S1 addresses the issue of inconsistent printhead temperature and motor current quality and dynamic changes in distribution over time in complex industrial environments. Firstly, it addresses the problem of inconsistent printhead temperature and motor current quality and distribution in thermal printers. An adaptive hybrid interpolation strategy intelligently handles missing values ​​caused by frequent start-stop or interference. Subsequently, adaptive phased data standardization is performed based on the thermal printer's runtime, providing high-quality, cross-phase comparable standardized feature sequences for the improved AGA-GRU-Informer model input. This fundamentally improves the reliability and robustness of data processing for this external device under real-world conditions. Step S2, targeting the complete lifecycle of the thermal printer from break-in, stabilization to aging, constructs a dynamic feature set that accurately reflects performance degradation at different stages. Through dynamic selection of health factors on the standardized sequences, features are automatically filtered out. The key influencing factors at each lifecycle stage were identified, and adaptive data partitioning was performed accordingly. This achieved a shift from static global features to staged and refined feature modeling that conforms to the actual degradation patterns of thermal printers. This allows the subsequently improved AGA-GRU-Informer model to adaptively diagnose and focus on the core degradation indicators of the current lifecycle stage. Step S3, to meet the stringent requirements for prediction accuracy and timeliness in industrial scenarios, constructs a dedicated improved AGA-GRU-Informer model. Based on the partitioned dataset, the improved AGA-GRU-Informer model is used. This model... The adaptive genetic algorithm optimizes the structure and integrates the advantages of GRU in capturing short-term operational fluctuations of thermal printers and Informer in capturing long-term performance degradation trends. This enables earlier and more accurate assessment and trend prediction of the device's health status. Through this progressive industrial-grade technology chain, from data processing to feature engineering to intelligent modeling, various challenges from complex data acquisition to high-reliability status assessment of thermal printers are solved. Ultimately, a thermal printer quality assessment and predictive maintenance method that is deeply adapted to key business scenarios, highly adaptable, highly accurate, and interpretable is achieved, effectively preventing unplanned downtime and demonstrating its engineering application value.

[0104] It should be noted that this invention is specifically adapted and innovated to address the particular industrial technical problem of health assessment throughout the entire life cycle of thermal printers. Specifically, the adaptive hybrid interpolation strategy does not select linear or quadratic interpolation in isolation, but is based on comparing the variance of the local data sequence with a preset threshold. This threshold is dynamically determined according to the fluctuation characteristics of the thermal printer under normal operating conditions (1.5 times the interquartile range), which solves the engineering problem of different data loss types in thermal printers under frequent start-stop and interference environments.

[0105] The adaptive phased standardization divides the life cycle of thermal printers into three stages for the first time: new machine stage (less than 500 hours), stable stage (500 to 3000 hours), and aging stage (greater than or equal to 3000 hours). The mean and standard deviation of each stage are calculated independently, which fundamentally eliminates the impact of data distribution drift throughout the equipment's life cycle on the evaluation model. The threshold for this division is based on the typical critical time determined by Weibull statistical analysis and has a clear physical meaning.

[0106] The dynamic selection of health factors automatically filters key monitoring parameters and assigns weights based on the correlation coefficients between different stages and health status (Pearson correlation coefficient for continuous data and point-to-bicollinear correlation coefficient for binary data). This solves the fundamental defect that traditional static feature sets cannot adapt to the dynamic changes of key factors during equipment aging. The first stage focuses on printing speed, temperature, and current; the second stage focuses on temperature, current, and paper condition; and the third stage focuses on current, running time, and temperature.

[0107] The improved AGA-GRU-Informer model uses an adaptive genetic algorithm to automatically optimize the GRU hidden layer dimension (ranging from 32 to 256), replacing manual parameter tuning, enabling the AGA-GRU-Informer model to adapt to the data characteristics of different devices; the GRU layer is deeply integrated with the Informer's ProbSparse self-attention mechanism: after the GRU extracts short-term dynamic features, its hidden state sequence is input into the Informer's sparse attention layer to capture long-term dependencies; the generative decoder takes long-term dependency features and future position encoding as input and outputs a multi-step health state prediction sequence;

[0108] This invention organically combines short-term fluctuation capture with long-term degradation trend modeling, thereby improving prediction accuracy compared to using GRU or Informer alone.

[0109] Step S1 includes steps S101, S102 and S103;

[0110] Step S101: Obtain the time-series data collected in chronological order during the operation of the thermal printer, as well as the corresponding maintenance and fault record data;

[0111] Time-series data includes monitoring data and attribute data;

[0112] The monitoring data includes printhead temperature data, motor current data, printing speed data, and paper transport status data.

[0113] The attribute data includes the thermal printer's runtime data and the thermal printer's ID;

[0114] Specifically, the paper transfer status data includes 0 for normal and 1 for paper jam.

[0115] Step S102: For time series data, the timestamp of the collected time series data is combined with the ID of the thermal printer as a unique identifier to identify and delete duplicate records, thereby obtaining deduplicated time series data.

[0116] If there are missing values ​​in the deduplicated time series data, an adaptive hybrid interpolation strategy is used to interpolate the missing values ​​to obtain the time series data matrix.

[0117] The adaptive hybrid interpolation strategy specifically includes selecting 10 samples before and after each data point with missing values ​​in the deduplicated time series data to form a local data sequence, calculating the variance of the local data sequence, and obtaining the local data sequence variance.

[0118] The variance of the local data sequence is compared with a preset missing threshold. If the variance of the local data sequence is less than the missing threshold, linear interpolation is used to interpolate the missing values.

[0119] If the variance of a local data sequence is greater than or equal to the missing threshold, then quadratic polynomial interpolation is used to interpolate the missing values.

[0120] After deduplication and interpolation, the time series data are arranged in the order of the timestamps of the collected time series data to obtain a time series data matrix.

[0121] Time series data matrix representation is ;

[0122] in, Indicates the number of samples. Indicates the quantity of monitoring data. This represents the timestamp, thermal printer ID, and runtime data of the time series data after deduplication and interpolation.

[0123] Step S103, adaptive phased data standardization processing specifically includes dividing the thermal printer's running time data in the time series data matrix according to interval division logic to obtain the corresponding life cycle stages;

[0124] The life cycle stages include the first stage, the second stage, and the third stage;

[0125] The interval division logic specifically includes dividing the data into the first stage when the running time is less than the first running threshold.

[0126] When the runtime data is greater than or equal to the first runtime threshold and less than the second runtime threshold, it is classified into the second stage;

[0127] When the runtime data is greater than or equal to the second runtime threshold, it is divided into the third stage;

[0128] Calculate the mean and standard deviation of all monitoring data in the first, second and third stages respectively;

[0129] For each data point in the time series data matrix, the corresponding life cycle stage is determined based on the corresponding running time data, and the standardization calculation is performed based on the mean and standard deviation corresponding to the life cycle stage. The calculation logic is as follows: the difference between the monitoring data of the data point and the corresponding mean is calculated to obtain the deviation value, and the ratio of the deviation value to the corresponding standard deviation is calculated to obtain the standardized monitoring data.

[0130] And set a standard range, and replace the monitoring data that exceeds the standard range with the boundary value corresponding to the standard range;

[0131] The processed standardized monitoring data, the timestamps of the deduplicated and interpolated time series data, the thermal printer ID, and the running time data are combined to form a standardized feature matrix;

[0132] The standardized feature matrix is ​​arranged in the order of the timestamps of the time series data after deduplication and interpolation to form a standardized feature sequence.

[0133] In one embodiment, step S101 provides a complete and accurate raw data foundation for the entire evaluation process, ensuring the primary prerequisite for the reliability of subsequent analysis. Specifically, step S101 involves acquiring time-series data and corresponding maintenance and fault records collected sequentially during the operation of the thermal printer. The time-series data includes monitoring data and attribute data. The monitoring data includes printhead temperature data (in degrees Celsius), motor current data (in amperes), printing speed data (in millimeters per second), and paper transfer status data. The attribute data includes the thermal printer's cumulative operating time data (in hours) and the thermal printer's unique identifier ID. The paper transfer status data is binary data, with 0 indicating normal operation and 1 indicating a paper jam. This step, by integrating multi-source information, provides the necessary device status labels and feature data for constructing a supervised learning framework.

[0134] Step S102 addresses the common issues of duplication and missing data in industrial field data by providing a robust and adaptive cleaning and repair method to improve the quality and consistency of time-series data. Specifically, Step S102 involves: first, combining the timestamp of the collected time-series data with the ID of the thermal printer as a unique identifier; identifying and deleting duplicate records with identical timestamp and ID combinations to obtain deduplicated time-series data; then, if missing values ​​exist in the deduplicated time-series data, an adaptive hybrid interpolation strategy is used to interpolate these missing values; this adaptive hybrid interpolation strategy specifically includes: for each data point with missing values, attempting to select 10 valid samples at each of its preceding and following time-series positions. Here, a valid sample refers to a data point where the monitored data is not null at that time-series position. If there are fewer than 10 valid samples before or after, all usable valid samples are used to form a local data sequence; calculating the variance of this local data sequence; and then comparing the local data sequence variance with a preset missing value threshold. The missing data threshold is set based on the fluctuation characteristics under normal operating conditions in historical data. Specifically, it includes calculating the difference between the third quartile and the first quartile of the variance of all monitoring data sequences during normal, fault-free operation, called the interquartile range, and using 1.5 times this interquartile range as the missing data threshold. If the variance of a local data sequence is less than this missing data threshold, the data is considered to have a gentle fluctuation, and linear interpolation is used to interpolate the missing values. If the variance of a local data sequence is greater than or equal to this missing data threshold, the data is considered to have a drastic fluctuation, and quadratic polynomial interpolation, which better fits the trend change, is used to interpolate the missing values. As a robustness guarantee, if the total number of valid samples in the constructed local data sequence is less than 3, local interpolation is abandoned, and the arithmetic mean of all non-missing values ​​of the monitoring data in the entire dataset, i.e., the global mean, is used for filling. Finally, all time series data after deduplication and interpolation are arranged strictly according to the chronological order of the timestamps of the collected time series data to obtain a regular time series data matrix. The dimension of this time series data matrix is ​​represented as follows: Multiply Add 3 columns, among which, This indicates the total number of samples. This indicates the number of types of monitored data, which is 4 here: printhead temperature data, motor current data, printing speed data, and paper transport status data. Adding 3 indicates the inclusion of 3 additional columns of attribute data: timestamp of the time-series data collection, thermal printer ID, and runtime data. By adaptively selecting the interpolation method based on local volatility, it can maintain simplicity and efficiency when the data is stable, and better fit the local trend when the data fluctuates. At the same time, the fallback strategy of global mean filling ensures the feasibility of handling extreme cases, thereby improving data integrity and providing high-quality input for subsequent analysis.

[0135] Step S103 addresses the issue of data distribution drift throughout the entire lifecycle of external devices. Phased standardization ensures comparability of monitoring data from different operating periods, which is crucial for accurate cross-cycle health status assessment. Specifically, Step S103 involves adaptive phased data standardization. This includes dividing the thermal printer's operating time data in the time-series data matrix according to a preset interval division logic to obtain the lifecycle stage corresponding to each data point. The lifecycle stages include the first stage, the second stage, and the third stage, corresponding to the new device's break-in period, stable operation period, and performance aging period, respectively. The interval division logic involves setting two operating time thresholds: the first threshold is set to 500 hours, based on statistical analysis of historical operating data from similar thermal printer groups. Given the fluctuation errors that occur when thermal printers are first connected to a computer, the first threshold ensures that the thermal printers enter a stable working state, improving the accuracy of subsequent steps S2 and S3. The second threshold is set to 3000 hours, based on the entire lifecycle operating data of this external device group. Based on maintenance records and Weibull statistical analysis, typical critical times from stable operating state to accelerated performance aging are determined. When the operating time data is less than a first operating threshold, it is classified into the first stage; when the operating time data is greater than or equal to the first operating threshold and less than a second operating threshold, it is classified into the second stage; when the operating time data is greater than or equal to the second operating threshold, it is classified into the third stage. Then, the arithmetic mean and standard deviation of each type of monitoring data are calculated independently for each of the first, second, and third stages. Next, for each data point in the time series data matrix, its life cycle stage is determined based on its corresponding operating time data. Standardization is then performed based on the mean and standard deviation of the corresponding monitoring data calculated for that life cycle stage. The calculation logic is as follows: subtract the mean of the stage from the monitoring data value of the data point to obtain the deviation value, and then divide the deviation value by the standard deviation of the stage to obtain the standardized monitoring data value. To ensure the robustness of the standardized data distribution and suppress the influence of extreme outliers, a standard interval based on statistical experience is set, for example, an interval from -3 to +3, which corresponds to approximately 99% of a normal distribution.For a 7% data range, monitoring data values ​​exceeding this range after standardization are uniformly replaced with the corresponding boundary values ​​of the range; that is, values ​​less than -3 are set to -3, and values ​​greater than +3 are set to +3. Finally, the processed standardized monitoring data, the timestamps of the deduplicated and interpolated time-series data, the thermal printer ID, and the runtime data are combined to form a standardized feature matrix. This feature matrix is ​​then arranged in timestamp order to form the final standardized feature sequence. Through phased standardization, the overall data distribution shift caused by equipment wear and performance degradation is eliminated, enabling the improved AGA-GRU-Informer model to more fairly compare the status of external equipment at different lifecycle stages. Simultaneously, tail reduction controls the negative impact of outliers, enhancing the overall stability of the data.

[0136] Step S1 constitutes the data preprocessing layer. Its core objective is to transform the raw, messy, missing, and unevenly distributed industrial time-series data into a high-quality, standardized, and time-comparable standard feature sequence. Through cleaning, repair, and standardization operations, the reliability and consistency of the input information are ensured, providing crucial data support for the subsequent accurate assessment of the health status of the thermal printer.

[0137] Step S2 includes steps S201 and S202;

[0138] Step S201: Extract maintenance events and their corresponding timestamps from the maintenance records of the thermal printer and fault events and their corresponding timestamps from the fault history data. Align the timestamps corresponding to the maintenance events and fault events with the timestamps of the time-series data after deduplication and interpolation in the standardized feature sequence.

[0139] For each data point in the standardized feature sequence, the health status value of the thermal printer corresponding to each data point in the standardized feature sequence is calculated by linear interpolation between adjacent maintenance events and fault events before and after the data point. This health status value is used as the actual health value. The range of the health status value is 0 to 1, where 1 indicates that the thermal printer is completely healthy and 0 indicates that the thermal printer is completely unresponsive.

[0140] The calculated true health status value of each data point is arranged according to the timestamp order of the time series data after deduplication and interpolation in the corresponding standardized feature sequence to form a true health status value sequence.

[0141] Based on interval partitioning logic, the corresponding life cycle stages are obtained. The monitoring data in the standardized feature sequence are grouped according to the life cycle stages, and the correlation coefficient between each monitoring data and the real health status value sequence in each life cycle stage is calculated. Specifically, the Pearson correlation coefficient between the printhead temperature data, motor current data, printing speed data, and running time data and the real health status value is calculated, and the point-to-binary correlation coefficient between the paper transfer status data and the real health status value is calculated.

[0142] For the first, second, and third stages of the life cycle, data that meets the preset screening criteria are selected from the monitoring data as health factors for the corresponding life cycle stage.

[0143] For the first stage, the specific screening criteria include that the absolute value of the correlation coefficient is greater than the first correlation threshold, and the monitored data belongs to printing speed data, printhead temperature data, or motor current data.

[0144] For the second stage, the specific screening criteria include that the absolute value of the correlation coefficient is greater than the first correlation threshold, and the monitored data belongs to the temperature data of the print head, the current data of the motor, or the paper transfer status data.

[0145] For the third stage, the specific screening criteria include that the absolute value of the correlation coefficient is greater than the first correlation threshold, and the monitored data belongs to motor current data, running time data, or printhead temperature data.

[0146] Based on the absolute value of the correlation coefficient of each health factor in the corresponding life cycle stage, a weight is assigned to each health factor. The initial weight of the health factor is equal to the absolute value of the corresponding correlation coefficient. The weights of all health factors in the same life cycle stage are normalized to obtain and output the first stage health factor set and corresponding weights, the second stage health factor set and corresponding weights, and the third stage health factor set and corresponding weights.

[0147] The first-stage health factor set, the second-stage health factor set, and the third-stage health factor set constitute the health factor dataset.

[0148] Step S202: Adaptively partition the health factor dataset to obtain the partitioned dataset, specifically the first dataset and the second dataset.

[0149] The samples in the health factor dataset are sorted in ascending order by the timestamps of the time series data after deduplication and interpolation, and then sorted according to the corresponding life cycle stage to obtain the first stage data sequence, the second stage data sequence, and the third stage data sequence.

[0150] Adaptive data partitioning specifically includes using a single decision strategy to divide all thermal printers into independent and non-overlapping datasets;

[0151] Based on the data from the first, second, and third phases, a corresponding sliding window is set for each lifecycle phase, including the new machine window, the stable period window, and the aging period window.

[0152] The first phase corresponds to the new opportunity window, with a window length of 100 samples;

[0153] The second phase corresponds to the stable period window, with a window length of 300 samples;

[0154] The third stage corresponds to the aging period window, with a window length of 500 samples;

[0155] A corresponding sliding window is generated for the first stage data sequence, the second stage data sequence, and the third stage data sequence. When generating the sliding window, the overlap ratio between adjacent windows is set as the window overlap threshold.

[0156] All thermal printers are divided into a first dataset and a second dataset according to a preset ratio;

[0157] The first dataset includes a health factor dataset of a first proportion, and the second dataset includes a health factor dataset of a second proportion.

[0158] In one embodiment, step S201 specifically includes: extracting maintenance events and their corresponding timestamps from the maintenance records of the thermal printer and fault events and their corresponding timestamps from the fault history data; aligning the timestamps of the maintenance events and fault events with the timestamps in the standardized feature sequence; for each data point in the standardized feature sequence, calculating the thermal printer health status value corresponding to each data point by performing linear interpolation between adjacent maintenance events and fault events before and after that data point, which serves as the true health status value. The true health status value ranges from 0 to 1, where 1 represents that the thermal printer is fully healthy and 0 represents that the thermal printer is completely unresponsive; for example, the health status recovers to 1 after one maintenance and drops to 0 when the next fault occurs, with the status value decreasing linearly over time at any point between the two points; and then, according to the time sequence of the calculated true health status value of each data point after deduplication and interpolation in the standardized feature sequence. The data are arranged in chronological order by timestamp to form a sequence of true health status values. Based on the interval division logic defined in step S103, the monitoring data in the standardized feature sequence are grouped according to life cycle stages, and the correlation coefficient between each monitoring data and the true health status value sequence is calculated for each life cycle stage. Specifically, the Pearson correlation coefficients of printhead temperature data, motor current data, printing speed data, and running time data with the true health status values ​​are calculated, and the point-to-point correlation coefficients of paper transfer status data with the true health status values ​​are calculated. For the first, second, and third stages, health factors corresponding to the stage are selected from the monitoring data according to preset screening conditions. The first correlation threshold is set to 0.3. The specific method for setting this threshold is to summarize and calculate the absolute values ​​of the correlation coefficients of all monitoring data with the true health status value sequence in all life cycle stages, and use the 75th percentile of this absolute value set to 0.3 is recommended as the first correlation threshold. For the first stage, the screening criteria are that the absolute value of the correlation coefficient is greater than this first correlation threshold, and the monitored data belongs to printing speed data, printhead temperature data, or motor current data. For the second stage, the screening criteria are that the absolute value of the correlation coefficient is greater than this first correlation threshold, and the monitored data belongs to printhead temperature data, motor current data, or paper transport status data. For the third stage, the screening criteria are that the absolute value of the correlation coefficient is greater than this first correlation threshold, and the monitored data belongs to motor current data, running time data, or printhead temperature data. Based on the absolute value of the correlation coefficient of each health factor in its corresponding life cycle stage, each health factor is considered as a health factor. Initial weights are assigned to health factors, equal to the absolute value of their correlation coefficients. Subsequently, the weights of all health factors within the same life cycle stage are normalized to a sum of 1, resulting in and outputting the first-stage health factor set and its corresponding weights, the second-stage health factor set and its corresponding weights, and the third-stage health factor set and its corresponding weights. These three sets together constitute the health factor dataset. Continuous monitoring signals are generated through maintenance and fault records, and based on a data-driven approach, the set of features most sensitive to health status at different life cycle stages and their weights are automatically identified, achieving dynamic and refined selection of health factors.

[0159] Step S202 specifically involves adaptively partitioning the health factor dataset to obtain the partitioned datasets, specifically the first dataset and the second dataset. First, the samples in the health factor dataset are sorted in ascending order by the timestamps of the time-series data after deduplication and interpolation, and then arranged according to the corresponding life cycle stages, resulting in the first-stage data sequence, the second-stage data sequence, and the third-stage data sequence. The adaptive data partitioning specifically includes the following operations: using a single decision strategy, based on the thermal printer ID, all thermal printer data are divided into independent and non-overlapping datasets; that is, all data from the same printer appears only in either the first dataset or the second dataset, to ensure improvement. To ensure the fairness of the AGA-GRU-Informer model evaluation, a corresponding sliding window was set for the data in each of the three phases: Phase 1, Phase 2, and Phase 3. The window length was determined based on the timescale of the external device's performance evolution at each stage of its lifecycle. Phase 1 corresponds to the new device period, with a window length of 100 samples, designed to cover the typical short cycle from initial operation to basic parameter stability. Phase 2 corresponds to the stable period, with a window length of 300 samples. Phase 3 corresponds to the aging period, with a window length of 500 samples, designed to cover the typical longer process from performance stability to the appearance of obvious degradation signs. This window serves as the basis for evaluating the fairness of the AGA-GRU-Informer model evaluation at each stage of the lifecycle. The three-stage data sequences each generate corresponding sliding windows. When generating these sliding windows, the overlap ratio between adjacent windows, i.e., the window overlap threshold, is set to 50%. This 50% overlap ratio ensures sufficient temporal continuity of the generated time-series samples, while effectively increasing the number and diversity of samples available for subsequent analysis and modeling. This allows for more intensive and smoother sampling of the gradual state changes within each lifecycle stage, thus laying a better data foundation for capturing subtle health degradation. Finally, all thermal printer data is divided into a first dataset and a second dataset according to a preset ratio. Specifically, 70% of the thermal printer data is assigned to the first dataset, and the remaining 30% is assigned to the second dataset. The second dataset, with a first proportion set to 70%, provides a sufficient data foundation for the construction and parameter optimization of the improved AGA-GRU-Informer model, ensuring that the constructed evaluation system has sufficient statistical representativeness and stability. The second proportion set to 30% forms a completely independent validation set, used to objectively evaluate the performance and generalization ability of the final improved AGA-GRU-Informer model, ensuring the reliability and fairness of the evaluation conclusions. Through a phased, device-specific, sliding window-based partitioning method, input samples for the improved AGA-GRU-Informer model are generated that can reflect the long-term evolution pattern of the time series while ensuring the independence between different devices.

[0160] Step S2 constitutes the dynamic feature engineering and sample preparation layer. Its core objective is to automatically construct a model closely related to the actual health degradation process of external equipment by utilizing the high-quality standardized data provided in step S1. This layer combines external equipment maintenance knowledge, statistical correlation analysis, and time-series sample construction technology to achieve the key transformation from raw monitoring data to samples usable by the intelligent model.

[0161] Step S3 includes steps S301, S302, S303, S304 and S305;

[0162] Step S301: An adaptive genetic algorithm is used to automatically optimize the hidden layer dimension parameters of the GRU layer to obtain the optimal hidden layer dimension.

[0163] Step S301 includes steps S3011, S3012, S3013, S3014, S3015, S3016 and S3017;

[0164] Step S3011: Within the preset range of hidden layer dimensions, a set of candidate dimension values ​​is randomly generated to form the initial population;

[0165] Repeat steps S3012 to S3016 until the preset maximum number of iterations is reached to obtain the final generation of candidate dimension value population;

[0166] Step S3012: For each candidate dimension value in the current population, repeat steps S3013 to S3015 to obtain the corresponding fitness.

[0167] Step S3013: Construct a simplified GRU network, where the hidden layer dimensions of the simplified GRU network are candidate dimension values;

[0168] Step S3014: Sort all samples in the first dataset in ascending order according to the timestamps of the time series data after deduplication and interpolation, and select the third proportion of samples with the earliest timestamps in the first dataset to form the evaluation data sequence.

[0169] The evaluation data sequence is divided into continuous time-series samples using a fixed-length sliding window;

[0170] Step S3015: Input the time series samples into a simplified GRU network for processing to obtain the predicted health status sequence;

[0171] Calculate the root mean square of the difference between the predicted health status sequence and the corresponding true health status value sequence, and use it as the root mean square error of the candidate dimension value.

[0172] And calculate the reciprocal of the root mean square error as the fitness of the current candidate dimension value;

[0173] Step S3016: Based on the fitness of all candidate dimension values ​​in the current population, perform selection, crossover, and mutation operations to generate a new generation of candidate dimension value population;

[0174] Step S3017: From the final generation of candidate dimension values, output the candidate dimension value with the highest fitness as the optimal hidden layer dimension.

[0175] The selection process specifically includes sorting all candidate dimension values ​​according to their fitness and retaining the fourth-highest fitness candidate dimension values ​​as the parent individuals.

[0176] The crossover specifically involves randomly selecting two candidate dimension values ​​from the parent individuals, denoted as... and The offspring dimension values ​​are generated through linear combination, expressed as follows:

[0177] ;

[0178] in, Represents the child dimension value. Represents a random number in the range [0,1].

[0179] The mutation specifically involves adaptively perturbing the generated child dimension values ​​to obtain the final child dimension values, expressed as follows:

[0180] ;

[0181] in, This represents the final child dimension value. For adaptive coefficient of variation, This represents normally distributed noise. This indicates that the mean is 0 and the variance is 0. The normally distributed noise, after variation, will Constraints Within the range.

[0182] Step S302: Extract feature sequence fragments from the first dataset and the second dataset according to the life cycle stage and the corresponding sliding window of the sample.

[0183] For each feature sequence segment, extract a continuous sequence segment with a preset step length following the feature sequence segment from the corresponding real health status value sequence, and use the continuous sequence segment as the target sequence of the feature sequence segment;

[0184] Based on the feature sequence fragments and the corresponding target sequences, the first time series sample and the second time series sample are generated respectively.

[0185] The process of generating time-series samples includes adding sine-cosine position codes to the feature sequence segments and the corresponding target sequences, respectively.

[0186] The first and second time series samples are processed by a gated recurrent unit layer to obtain the corresponding GRU hidden state sequences.

[0187] Step S303: Normalize each GRU hidden state vector in the GRU hidden state sequence to obtain a normalized feature sequence.

[0188] The improved AGA-GRU-Informer model employs three independent linear transformation layers to map the normalized feature sequence into a query matrix, a key matrix, and a value matrix, respectively.

[0189] Based on the improved AGA-GRU-Informer model, the ProbSparse self-attention mechanism uses the query matrix and key matrix to filter queries and obtain a simplified query matrix. Specifically, for each query vector in the query matrix, the similarity distribution between the query vector and the set of all key vectors in the key matrix is ​​calculated.

[0190] Based on the similarity distribution, calculate the sparsity score for each query vector;

[0191] Based on sparsity scores, the top performers are selected from the query matrix. A number of important query vectors are used to form a simplified query matrix, where u is the preset number of important queries;

[0192] Step S304, simplify the query matrix Key matrix Sum matrix The input is fed into the ProbSparse attention computation layer of the improved AGA-GRU-Informer model, and the long-term dependency feature representation is obtained by computation using the standard dot product attention formula.

[0193] Step S305: The long-term dependency feature representation and the sine-cosine position encoding of the target sequence are input into the generative decoder of the improved AGA-GRU-Informer model. The generative decoder processes the data through a three-layer self-attention mechanism and a fully connected layer, and outputs the context aggregation representation.

[0194] The context aggregation representation is processed by linear transformation and activation function to output a predicted sequence of the health status of the thermal printer.

[0195] Based on the mapping of context aggregation representation, the predicted health status value of thermal printer is obtained. Specifically, it includes extracting the feature vector of the starting position of the target sequence in the context aggregation representation, and mapping the feature vector to an unscaled original prediction value through a fully connected layer.

[0196] The original predicted values ​​are processed by the Sigmoid activation function to obtain the final predicted health status values.

[0197] The predicted value ranges from 0 to 1, where 1 indicates that the thermal printer is fully healthy and 0 indicates that it is completely unusable.

[0198] In one embodiment, step S3 involves constructing and applying an improved AGA-GRU-Informer model that integrates short-term dynamic capture and long-term dependency modeling capabilities to achieve accurate assessment and prediction of the health status of the thermal printer. Step S3 specifically includes step S301, which automatically determines the optimal structural parameters of the gated recurrent unit network layer to replace tedious manual parameter tuning. This pre-optimization step enhances the basic feature extraction capability and algorithm adaptability of the improved AGA-GRU-Informer model. Step S301 specifically involves using an adaptive genetic algorithm to automatically optimize the hidden layer dimension parameters of the GRU layer to obtain the optimal hidden layer dimension. This includes the following sub-steps: Step S... 3011, within a preset range of hidden layer dimensions, between 32 and 256 based on common network sizes and computational resources. This range is designed to cover typical configurations from lightweight to more complex models, providing ample search space for the optimization algorithm. For example, an intermediate value such as 128 can be selected as an exemplary starting reference point, but the final optimal value will be automatically determined by an adaptive genetic algorithm within the preset range. A set of 50 candidate dimension values ​​is randomly generated to form the initial population. These 50 candidate dimension values ​​provide a sufficiently diverse initial search population for the adaptive genetic algorithm based on evolutionary principles, effectively exploring the solution space to find potential optimal solutions while avoiding excessive population size. The large size of the algorithm leads to excessive computational overhead in the initial optimization process, thus achieving a good balance between optimization effect and computational efficiency. Steps S3012 to S3016 constitute a complete iteration, which is repeated until the preset maximum number of iterations is reached. The maximum number of iterations is set to 50, resulting in the final generation of candidate dimension value population. 50 iterations provide sufficient generations to ensure that superior genes have enough opportunities to be preserved and combined, thereby stably approaching the optimal solution region. At the same time, this setting can effectively prevent the algorithm from getting stuck in excessive meaningless iterations, avoiding the waste of computational resources. Finally, a reasonable engineering balance is achieved between the sufficiency of the optimization process and the overall computational efficiency. Step S3012, for each of the current population... The evaluation process is performed on the candidate dimension values. Step S3013: A simplified GRU network is constructed, which includes only one GRU layer and one output layer. The hidden layer dimension is the candidate dimension value to be evaluated. Step S3014: From the first dataset, the earliest 20% of the samples are selected to form the evaluation data sequence. This proportion is set to ensure that the evaluation data sequence includes a sufficient number of samples for stable performance statistics, while minimizing the overall computational cost of the automated optimization process. A fixed-length sliding window is used to divide the evaluation data sequence into continuous time-series samples. The window length refers to the window length of the corresponding life cycle stage defined in step S202, i.e., 100 samples in the new machine period window.Step S3015: Input the time-series sample into the simplified GRU network for forward propagation to obtain the predicted health state sequence; calculate the root mean square error between the predicted sequence and the corresponding real health state value sequence; calculate the reciprocal of the root mean square error as the fitness of the current candidate dimension value; Step S3016: Perform selection, crossover, and mutation operations based on the population fitness to generate a new generation population; the selection operation is to sort according to fitness and retain the top 20% of individuals with the highest fitness as parents; the crossover operation is to randomly select two candidate dimension values ​​from the parent individuals, denoted as ; and The algorithm generates offspring dimension values ​​through linear combination. The mutation operation involves adaptively perturbing the generated offspring dimension values ​​to obtain the final offspring dimension values. The adaptive mutation coefficient is initially set to 0.1, and σ² represents the noise variance, set to 1 to ensure effective exploration. The final offspring dimension values ​​are constrained within a preset dimension range, between 32 and 256. This ensures that all candidate network structures generated by the evolutionary algorithm are within a reasonably complex complexity range. The lower limit of 32 prevents insufficient network expressive power due to too low a hidden layer dimension, thus avoiding ineffective capture of healthy state features. The upper limit of 256 avoids redundant model parameters due to excessively high dimensions, preventing overfitting and unnecessary consumption of computational resources. By strictly constraining the search space... Within this scope, the adaptive genetic algorithm can concentrate computational resources on efficient optimization within the effective solution space while ensuring the potential performance of the improved AGA-GRU-Informer model. This ensures that the final optimal hidden layer dimension has both sufficient capacity for the improved AGA-GRU-Informer model and the feasibility and efficiency of practical deployment. In step S3017, the candidate dimension value with the highest fitness is output from the final generation population as the optimal hidden layer dimension. Through evolutionary calculation, the GRU network structure parameters adapted to the current thermal printer data characteristics are automatically obtained, laying an efficient and customized feature extraction foundation for the subsequent improved AGA-GRU-Informer model.

[0199] Step S302 transforms the original health factor data and target values ​​into time-series sample pairs containing location information that can be processed by the neural network. A GRU layer is then used to initially extract short-term dynamic features of the sequences. Specifically, Step S302 involves truncating the first and second datasets according to the life cycle stage corresponding to the samples and the corresponding sliding window length defined in Step S202, resulting in a series of feature sequence segments. For each feature sequence segment, a continuous sequence segment with a preset step size (e.g., 10 time units) immediately following the segment is extracted from the corresponding real health status value sequence as the target sequence. This target sequence provides a key future window length for timely and effective early warning of predictive maintenance. Based on the feature sequence fragments and the corresponding target sequences, first and second time-series samples are generated respectively. The generation process of time-series samples includes adding sine-cosine position codes to the feature sequence fragments and the target sequences respectively to inject the sequence order information. Subsequently, the first and second time-series samples are input into a gated recurrent unit (GRU) layer for processing. The hidden layer dimension of this GRU layer adopts the optimal hidden layer dimension obtained in step S301, thereby obtaining the corresponding GRU hidden state sequence. This completes the transformation from structured data to deep learning samples and preliminarily encodes the temporal dependencies of the input sequence through the GRU layer, providing a basic representation for capturing more complex patterns in the future.

[0200] Step S303, based on the short-term features extracted by GRU, introduces an efficient sparse attention mechanism to focus on key information, thereby capturing long-term dependencies affecting health status. Specifically, Step S303 involves: first, performing layer normalization on each vector in the GRU hidden state sequence to obtain a normalized feature sequence; then, using three independent linear transformation layers to map the normalized feature sequence to a query matrix Q, a key matrix K, and a value matrix V, respectively; next, based on the improved AGA-GRU-Informer model's ProbSparse self-attention mechanism, using the query matrix and key matrix for query filtering to obtain a simplified query matrix. Specifically, for each query vector in the query matrix, the similarity distribution between the query vector and the set of all key vectors in the key matrix is ​​calculated; based on the similarity distribution, according to the sparsity measurement method defined by the ProbSparse self-attention mechanism, the sparsity score of each query vector is calculated; based on the sparsity score, the top u important query vectors are selected from the query matrix to form a simplified query matrix, where u is a preset number of important queries. The important query quantity *u* is an adjustable hyperparameter. Its specific value is determined as follows: based on the first time-series sample generated in step S302, 10% of the samples are reserved as a validation subset. This provides an independent and stable evaluation benchmark for optimizing the hyperparameter *u*, ensuring that the selected *u* value enables the improved AGA-GRU-Informer model to achieve good generalization performance on unseen data. Simultaneously, this proportion balances the statistical reliability of the validation with the sufficiency of data for constructing the AGA-GRU-Informer model used for main process improvement. Within a preset range of candidate *u* values, the prediction performance of the improved AGA-GRU-Informer model corresponding to different *u* values ​​is evaluated on this validation subset, and the *u* value that minimizes the prediction error is selected as the final setting. The candidate range can be based on the total query length. Settings, for example, from 1 to Integer search is performed between them; by filtering important queries, the computational complexity of attention is greatly reduced, and the improved AGA-GRU-Informer model can focus on the most informative part of the sequence, which is ready for efficiently capturing long-term dependencies;

[0201] Step S304 involves performing attention calculations using the filtered key queries and complete key-value pairs to fuse long-term contextual information and form rich feature representations. Specifically, step S304 involves refining the query matrix... The key matrix K and value matrix V are input into the ProbSparse attention computation layer, where they are calculated using the standard dot product attention formula to obtain the long-term dependency feature representation. The expression for the standard dot product attention formula is as follows: ;in, This represents the long-term dependency feature representation. This represents the filtered query matrix. Indicates the number of important queries. Represents the key matrix. This represents the transpose operator. Represents a value matrix, This represents the scaling factor; while ensuring computational efficiency, it aggregates the global context information of the sequence to generate high-level features that can characterize the long-term evolution trend of the device's health status.

[0202] Step S305: Based on long-term dependency features and positional information of future time steps, decode and generate a health status prediction sequence for a future period, and output specific health status prediction values. Specifically, step S305 involves inputting the long-term dependency feature representation and the sine-cosine positional encoding of the target sequence into a generative decoder. This decoder includes a network structure with a three-layer self-attention mechanism, a common configuration for this type of decoder, used to further aggregate contextual information; its output is a contextual aggregation representation. Subsequently, the contextual aggregation representation is processed through a linear transformation layer and a sigmoid activation function, finally outputting the health status prediction of the thermal printer within a preset future step size. The sequence has a preset step size that is exactly the same as the target sequence length defined in step S302, i.e., 10 time units. To obtain an instantaneous health status assessment value, the specific method is to extract the feature vector corresponding to the starting position of the target sequence in the context aggregation representation, map it to a scalar through a fully connected layer, and then activate it through the Sigmoid function to obtain the final health status prediction value with a value range between 0 and 1, where 1 represents complete health and 0 represents complete failure. This realizes the end-to-end mapping from historical sequences to future health status, which can output both trend prediction sequences and provide quantitative assessment values ​​at the current moment, thus meeting the dual needs of predictive maintenance and status assessment.

[0203] Step S3 constitutes the intelligent assessment and prediction layer. Its core objective is to integrate optimized sequence modeling units and efficient attention mechanisms to build a hybrid model that can simultaneously understand short-term operational fluctuations and long-term degradation trends of the equipment. This layer is the intelligent core of the entire method. It transforms the high-quality data processed in the preceding steps into profound insights and accurate predictions of the future health status of the equipment, ultimately achieving automation, accuracy, and foresight in the quality assessment of thermal printers.

[0204] The improved AGA-GRU-Informer model's structure, loss function, and training method specifically include:

[0205] The improved AGA-GRU-Informer model specifically includes the following structures:

[0206] Input layer: Receives the health factor dataset after processing in steps S1 and S2. Each sample is a two-dimensional tensor. The first dimension is the window length, and the second dimension is the feature dimension. The window length is set to 100 (newborn period), 300 (stable period), and 500 (aging period) according to the life cycle stage. The feature dimension is the number of health factors selected in each stage (e.g., 3 in the first stage, 3 in the second stage, and 3 in the third stage).

[0207] Position encoding layer: Sine-cosine position encoding is added to both the input sequence fragment and the target sequence (length is a preset step size, 10 time units in this embodiment) to inject sequence order information. The encoding method adopts the standard definition of Transformer;

[0208] Gated Recurrent Unit Layer: A single-layer GRU is adopted. The hidden layer dimension h is automatically optimized by an adaptive genetic algorithm within the range of 32 to 256 (see step S301 for the optimization process). The GRU processes the input sequence according to its standard gating mechanism (including update gate, reset gate, and candidate hidden state), and the output is a hidden state sequence with a sequence length equal to the window length.

[0209] Layer normalization: Perform layer normalization on each hidden state vector output by GRU to obtain a normalized feature sequence. The normalization operation adopts the standard definition (calculate the mean and variance independently for each sample and perform scaling and translation).

[0210] Linear transformation mapping: The normalized feature sequence is mapped to the query matrix, key matrix and value matrix respectively through three independent linear layers (without bias). The dimension of the mapped matrix is ​​equal to the dimension of the GRU hidden layer.

[0211] ProbSparse Self-Attention Layer: Following the ProbSparse self-attention mechanism of Informer, the sparsity score is first calculated for each query vector. Then, the top u query vectors with the highest scores are selected to form a simplified query matrix. The calculation method of u is: take the natural logarithm of the query length and then multiply it by the sampling factor (the sampling factor is 5 in this embodiment), and finally round up. Then, the standard dot product attention calculation formula is used to perform attention calculation with the simplified query matrix and the complete key matrix and value matrix to obtain the long-term dependency feature representation. This layer greatly reduces the computational complexity while preserving the long-term dependency information of the sequence.

[0212] Generative Decoder: The decoder uses a three-layer multi-head self-attention network (4 attention heads per layer) and a feedforward network. Its structure is the same as the generative decoder in the Informer model. The long-term dependency feature representation obtained in the previous step is concatenated with the positional encoding of the target sequence and then input into the decoder. The decoder outputs the context aggregation representation.

[0213] Output layer: Extract the feature vector corresponding to the starting position of the target sequence from the context aggregation representation, map the vector to a scalar through a fully connected layer, and then activate it through the Sigmoid function to obtain the final health status prediction value. The prediction value ranges from 0 to 1, where 1 represents complete health and 0 represents complete failure.

[0214] The loss function specifically includes:

[0215] This invention uses root mean square error as the loss function for training the improved AGA-GRU-Informer model. For multiple samples in a training batch, each sample includes a sequence of true health status values ​​(obtained by linear interpolation of maintenance records and fault records) and a corresponding prediction sequence (i.e., the predicted health status values ​​within a future preset step size output by the improved AGA-GRU-Informer model). The specific calculation method of the loss function is as follows: first, calculate the square of the difference between the predicted value and the true value at each time step, sum it over all samples and all time steps in the same batch, then divide it by the total number of prediction points, and finally take the square root. The smaller the value of this loss function, the higher the model prediction accuracy.

[0216] In step S301, the fitness function for the simplified GRU network (used only for hidden layer dimension optimization) is defined as the reciprocal of the root mean square error. That is, after calculating the root mean square error on the validation subset, its reciprocal is taken as the fitness. The higher the fitness, the better the candidate dimension value.

[0217] The improved training method for the AGA-GRU-Informer model specifically includes:

[0218] The training process uses the backpropagation algorithm and the Adam optimizer, with the specific hyperparameter settings as follows:

[0219] Optimizer: Adam, initial learning rate 0.001, weight decay coefficient 0.00001;

[0220] Learning rate scheduling: A cosine annealing decay strategy is adopted, with the minimum learning rate set to 0.000001 and the decay period being 50 training rounds.

[0221] Batch size: 64 samples per batch;

[0222] Maximum number of training epochs: 200 epochs, with an early stopping strategy: if the loss value on the validation set does not decrease for 20 consecutive epochs, training will be stopped early and the model parameters will be restored to the minimum value when the validation loss is minimized.

[0223] Validation set partitioning: 10% of the samples are randomly selected from the first dataset as the validation set, which is used for hyperparameter selection (such as the number of important queries u) and early stopping monitoring;

[0224] Gradient clipping: Set the upper limit of the gradient norm to 1.0 to prevent gradient explosion;

[0225] The training process specifically includes:

[0226] The optimal GRU hidden layer dimension, denoted as h*, is determined using an adaptive genetic algorithm (step S301);

[0227] An improved AGA-GRU-Informer model is constructed, including a GRU layer, a ProbSparse self-attention layer, and a generative decoder, and the hidden layer dimension of the GRU layer is fixed to h*.

[0228] The first dataset was divided into a training set (90%) and a validation set (10%), and the second dataset was used as an independent test set.

[0229] The training process is executed according to the above hyperparameter settings (including Adam optimizer, initial learning rate 0.001, weight decay 0.00001, cosine annealing decay, batch size 64, maximum number of training epochs 200, early stopping policy, gradient clipping limit 1.0). After each training epoch, the loss value on the validation set is calculated, and the model parameters with the minimum validation loss are saved.

[0230] After training, the performance of the final model is evaluated on the second dataset (test set), and metrics such as root mean square error, mean absolute error, and coefficient of determination are calculated.

[0231] Comparative experiments were conducted on three different batches of thermal printers. Specifically, a control group using a traditional fixed interpolation plus global normalization plus static feature set plus a single LSTM prediction model was compared with the proposed solution. The experimental results showed that the average root mean square error (RMSE) on the test set was 0.037, while that of the control group was 0.092, representing an improvement of approximately 60% in prediction accuracy. Throughout the entire lifecycle of the equipment, from its new phase to its aging phase, the prediction error fluctuation range of the improved AGA-GRU-Informer model did not exceed 0.01, while the error fluctuation of the control group exceeded 0.05. Compared to using GRU or Informer alone, the RMSE of the improved AGA-GRU-Informer model was reduced by 18% and 22% respectively, under the same number of training epochs. This solution addresses the technical challenges in thermal printer health assessment, including missing data, distribution drift, dynamic changes in key features, and the difficulty in integrating short-term fluctuations with long-term degradation.

[0232] Example 2, refer to Figure 2 In another embodiment of the present invention, which differs from the first embodiment, a computer peripheral device quality assessment system is provided, including a processing module, a partitioning module, and an assessment module;

[0233] The processing module acquires the time-series data of the thermal printer, uses an adaptive hybrid interpolation strategy to interpolate the missing values ​​of the time-series data to obtain a time-series data matrix, and performs adaptive phased data standardization processing on the time-series data matrix to obtain a standardized feature sequence.

[0234] The partitioning module performs dynamic selection of health factors on the standardized feature sequences to obtain a health factor dataset, and then performs adaptive data partitioning on the health factor dataset to obtain the partitioned dataset.

[0235] The evaluation module, based on the first and second datasets, uses an improved AGA-GRU-Informer model to evaluate and predict the health status of thermal printers.

[0236] In one embodiment, this progressive industrial-grade technology chain, from data processing to feature engineering to intelligent modeling, solves various challenges from complex data acquisition of thermal printers to high-reliability status assessment. Ultimately, it realizes a thermal printer quality assessment and predictive maintenance method that is deeply adapted to key business scenarios, highly adaptable, highly accurate and interpretable, effectively preventing unplanned downtime and demonstrating its engineering application value.

[0237] This invention effectively addresses the dynamic changes in data distribution throughout the entire lifecycle of equipment by employing an adaptive hybrid interpolation strategy and phased data standardization, significantly improving the quality and consistency of input data and laying a reliable foundation for subsequent analysis. The proposed dynamic health factor selection mechanism automatically filters and assigns weights to key monitoring parameters most relevant to the health status based on different operational stages of the equipment, solving the fundamental problem that traditional static feature sets cannot adapt to the equipment degradation process. The improved AGA-GRU-Informer hybrid model optimizes the network structure through an adaptive genetic algorithm and combines the ability of gated recurrent units to capture short-term dynamics with the ability of an Informer model integrating ProbSparse self-attention to capture long-term dependencies, achieving more accurate fitting and prediction of complex degradation trends. This solution forms a closed-loop optimization from data preprocessing and feature construction to the prediction model, ultimately significantly improving the accuracy, adaptability, and timeliness of thermal printer health status assessment and enhancing the model's generalization ability across different devices.

[0238] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0239] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A method for evaluating the quality of computer peripherals, characterized in that, Includes the following steps, Step S1: Obtain the time series data of the thermal printer, use an adaptive hybrid interpolation strategy to interpolate the missing values ​​of the time series data to obtain a time series data matrix, and perform adaptive phased data standardization processing on the time series data matrix to obtain a standardized feature sequence. Step S2: Perform dynamic selection of health factors on the standardized feature sequence to obtain a health factor dataset. Perform adaptive data partitioning on the health factor dataset to obtain the partitioned dataset. Step S3: Based on the first and second datasets, the improved AGA-GRU-Informer model is used to assess and predict the health status of the thermal printer.

2. The method for evaluating the quality of computer peripherals as described in claim 1, characterized in that, Step S1 includes steps S101, S102 and S103; Step S101: Obtain the time sequence data and corresponding maintenance and fault record data collected in chronological order during the operation of the thermal printer; Time-series data includes monitoring data and attribute data; The monitoring data includes printhead temperature data, motor current data, printing speed data, and paper transport status data. The attribute data includes the thermal printer's runtime data and the thermal printer's ID; Specifically, the paper transfer status data includes 0 for normal and 1 for paper jam. Step S102: For time series data, the timestamp of the collected time series data is combined with the ID of the thermal printer as a unique identifier to identify and delete duplicate records, thereby obtaining deduplicated time series data. If there are missing values ​​in the deduplicated time series data, an adaptive hybrid interpolation strategy is used to interpolate the missing values ​​to obtain the time series data matrix. The adaptive hybrid interpolation strategy specifically includes selecting 10 samples before and after each data point with missing values ​​in the deduplicated time series data to form a local data sequence, calculating the variance of the local data sequence, and obtaining the local data sequence variance. The variance of the local data sequence is compared with a preset missing threshold. If the variance of the local data sequence is less than the missing threshold, linear interpolation is used to interpolate the missing values. If the variance of a local data sequence is greater than or equal to the missing threshold, then quadratic polynomial interpolation is used to interpolate the missing values. After deduplication and interpolation, the time series data are arranged in the order of the timestamps of the collected time series data to obtain a time series data matrix. Time series data matrix representation is ; in, Indicates the number of samples. Indicates the quantity of monitoring data. The timestamp, thermal printer ID, and runtime data represent the time series data after deduplication and interpolation.

3. The method for evaluating the quality of computer peripherals as described in claim 2, characterized in that, Step S103, adaptive phased data standardization processing specifically includes dividing the thermal printer's running time data in the time series data matrix according to interval division logic to obtain the corresponding life cycle stages; The life cycle stages include the first stage, the second stage, and the third stage; The interval division logic specifically includes dividing the data into the first stage when the running time is less than the first running threshold. When the runtime data is greater than or equal to the first runtime threshold and less than the second runtime threshold, it is classified into the second stage; When the runtime data is greater than or equal to the second runtime threshold, it is divided into the third stage; Calculate the mean and standard deviation of all monitoring data in the first, second and third stages respectively; For each data point in the time series data matrix, the corresponding life cycle stage is determined based on the corresponding running time data, and the standardization calculation is performed based on the mean and standard deviation corresponding to the life cycle stage. The calculation logic is as follows: the difference between the monitoring data of the data point and the corresponding mean is calculated to obtain the deviation value, and the ratio of the deviation value to the corresponding standard deviation is calculated to obtain the standardized monitoring data. A standard range is set, and monitoring data exceeding the standard range is replaced with the boundary value corresponding to the standard range. The processed standardized monitoring data, the timestamps of the deduplicated and interpolated time series data, the thermal printer ID, and the running time data are combined to form a standardized feature matrix; The standardized feature matrix is ​​arranged in the order of the timestamps of the time series data after deduplication and interpolation to form a standardized feature sequence.

4. The method for evaluating the quality of computer peripherals as described in claim 3, characterized in that, Step S2 includes steps S201 and S202; Step S201: Extract maintenance events and their corresponding timestamps from the maintenance records of the thermal printer and fault events and their corresponding timestamps from the fault history data. Align the timestamps corresponding to the maintenance events and fault events with the timestamps of the time-series data after deduplication and interpolation in the standardized feature sequence. For each data point in the standardized feature sequence, the health status value of the thermal printer corresponding to each data point in the standardized feature sequence is calculated by linear interpolation between the maintenance events and fault events adjacent to the data point. This health status value is used as the actual health value. The range of the health status value is 0 to 1, where 1 indicates that the thermal printer is completely healthy and 0 indicates that the thermal printer is completely unresponsive. The calculated true health status value of each data point is arranged according to the timestamp order of the time series data after deduplication and interpolation in the corresponding standardized feature sequence to form a true health status value sequence. Based on interval partitioning logic, the corresponding life cycle stages are obtained. The monitoring data in the standardized feature sequence are grouped according to the life cycle stages, and the correlation coefficient between each monitoring data and the real health status value sequence in each life cycle stage is calculated. Specifically, the Pearson correlation coefficient between the printhead temperature data, motor current data, printing speed data, and running time data and the real health status value is calculated, and the point-to-binary correlation coefficient between the paper transfer status data and the real health status value is calculated.

5. The method for evaluating the quality of computer peripherals as described in claim 4, characterized in that, For the first, second, and third stages of the life cycle, data that meets the preset screening criteria are selected from the monitoring data as health factors for the corresponding life cycle stage. For the first stage, the specific screening criteria include that the absolute value of the correlation coefficient is greater than the first correlation threshold, and the monitored data belongs to printing speed data, printhead temperature data, or motor current data. For the second stage, the specific screening criteria include that the absolute value of the correlation coefficient is greater than the first correlation threshold, and the monitoring data belongs to the temperature data of the print head, the motor current data, or the paper transfer status data. For the third stage, the specific screening criteria include that the absolute value of the correlation coefficient is greater than the first correlation threshold, and that the monitoring data belongs to motor current data, running time data, or printhead temperature data. Based on the absolute value of the correlation coefficient of each health factor in the corresponding life cycle stage, a weight is assigned to each health factor. The initial weight of the health factor is equal to the absolute value of the corresponding correlation coefficient. The weights of all health factors in the same life cycle stage are normalized to obtain and output the first stage health factor set and corresponding weights, the second stage health factor set and corresponding weights, and the third stage health factor set and corresponding weights. The first-stage health factor set, the second-stage health factor set, and the third-stage health factor set constitute the health factor dataset.

6. The method for evaluating the quality of computer peripherals as described in claim 5, characterized in that, Step S202: Adaptively partition the health factor dataset to obtain the partitioned dataset, specifically the first dataset and the second dataset. The samples in the health factor dataset are sorted in ascending order by the timestamps of the time series data after deduplication and interpolation, and then sorted according to the corresponding life cycle stage to obtain the first stage data sequence, the second stage data sequence, and the third stage data sequence. Adaptive data partitioning specifically includes using a single decision strategy to divide all thermal printers into independent and non-overlapping datasets; Based on the data from the first, second, and third phases, a corresponding sliding window is set for each lifecycle phase, including the new machine window, the stable period window, and the aging period window. The first phase corresponds to the new opportunity window, with a window length of 100 samples; The second phase corresponds to the stable period window, with a window length of 300 samples; The third stage corresponds to the aging period window, with a window length of 500 samples; A corresponding sliding window is generated for the first stage data sequence, the second stage data sequence, and the third stage data sequence. When generating the sliding window, the overlap ratio between adjacent windows is set as the window overlap threshold. All thermal printers are divided into a first dataset and a second dataset according to a preset ratio; The first dataset includes a health factor dataset of a first proportion, and the second dataset includes a health factor dataset of a second proportion.

7. The method for evaluating the quality of computer peripherals as described in claim 6, characterized in that, Step S3 includes steps S301, S302, S303, S304 and S305; Step S301: An adaptive genetic algorithm is used to automatically optimize the hidden layer dimension parameters of the GRU layer to obtain the optimal hidden layer dimension. Step S301 includes steps S3011, S3012, S3013, S3014, S3015, S3016 and S3017; Step S3011: Within the preset range of hidden layer dimensions, a set of candidate dimension values ​​is randomly generated to form the initial population; Repeat steps S3012 to S3016 until the preset maximum number of iterations is reached to obtain the final generation of candidate dimension value population; Step S3012: For each candidate dimension value in the current population, repeat steps S3013 to S3015 to obtain the corresponding fitness. Step S3013: Construct a simplified GRU network, where the hidden layer dimensions of the simplified GRU network are candidate dimension values; Step S3014: Sort all samples in the first dataset in ascending order according to the timestamps of the time series data after deduplication and interpolation, and select the third proportion of samples with the earliest timestamps in the first dataset to form the evaluation data sequence. The evaluation data sequence is divided into continuous time-series samples using a fixed-length sliding window; Step S3015: Input the time series samples into a simplified GRU network for processing to obtain the predicted health status sequence; Calculate the root mean square of the difference between the predicted health status sequence and the corresponding true health status value sequence, and use it as the root mean square error of the candidate dimension value. The reciprocal of the root mean square error is calculated as the fitness of the current candidate dimension value; Step S3016: Based on the fitness of all candidate dimension values ​​in the current population, perform selection, crossover, and mutation operations to generate a new generation of candidate dimension value population; Step S3017: From the final generation of candidate dimension values, output the candidate dimension value with the highest fitness as the optimal hidden layer dimension. The selection process specifically includes sorting all candidate dimension values ​​according to their fitness and retaining the fourth-highest fitness candidate dimension values ​​as the parent individuals. The crossover specifically involves randomly selecting two candidate dimension values ​​from the parent individuals, denoted as... and The offspring dimension values ​​are generated through linear combination, expressed as follows: ； in, Represents the child dimension value. Represents a random number in the range [0,1]. The mutation specifically involves adaptively perturbing the generated child dimension values ​​to obtain the final child dimension values, expressed as follows: ； in, This represents the final child dimension value. For adaptive coefficient of variation, This represents normally distributed noise. This indicates that the mean is 0 and the variance is 0. The normally distributed noise, after variation, will Constraints Within the range.

8. The method for evaluating the quality of computer peripherals as described in claim 7, characterized in that, Step S302: Extract feature sequence fragments from the first dataset and the second dataset according to the life cycle stage and the corresponding sliding window of the sample. For each feature sequence segment, extract a continuous sequence segment with a preset step length following the feature sequence segment from the corresponding real health status value sequence, and use the continuous sequence segment as the target sequence of the feature sequence segment; Based on the feature sequence fragments and the corresponding target sequences, the first time series sample and the second time series sample are generated respectively. The process of generating the time-series samples includes adding sine-cosine position codes to the feature sequence segments and the corresponding target sequences, respectively. The first and second time series samples are processed by a gated recurrent unit layer to obtain the corresponding GRU hidden state sequences. Step S303: Normalize each GRU hidden state vector in the GRU hidden state sequence to obtain a normalized feature sequence. The improved AGA-GRU-Informer model employs three independent linear transformation layers to map the normalized feature sequence into a query matrix, a key matrix, and a value matrix, respectively. Based on the improved AGA-GRU-Informer model, the ProbSparse self-attention mechanism uses the query matrix and key matrix to filter queries and obtain a simplified query matrix. Specifically, for each query vector in the query matrix, the similarity distribution between the query vector and the set of all key vectors in the key matrix is ​​calculated. Based on the similarity distribution, the sparsity score of each query vector is calculated; Based on the sparsity score, the top [number] [items] are selected from the query matrix. A number of important query vectors are used to form a simplified query matrix, where u is the preset number of important queries; Step S304, simplify the query matrix Key matrix Sum matrix The input is fed into the ProbSparse attention computation layer of the improved AGA-GRU-Informer model, and the long-term dependency feature representation is obtained by computation using the standard dot product attention formula.

9. The method for evaluating the quality of computer peripherals as described in claim 8, characterized in that, Step S305: The long-term dependency feature representation and the sine-cosine position encoding of the target sequence are input into the generative decoder of the improved AGA-GRU-Informer model. The generative decoder processes the data through a three-layer self-attention mechanism and a fully connected layer, and outputs the context aggregation representation. The context aggregation representation is processed by linear transformation and activation function to output a predicted sequence of the health status of the thermal printer. Based on the mapping of context aggregation representation, the predicted health status value of the thermal printer is obtained. Specifically, this includes extracting the feature vector of the starting position of the target sequence in the context aggregation representation, and mapping the feature vector to an unscaled original prediction value through a fully connected layer. The original predicted value is input into the Sigmoid activation function for processing to obtain the final health status prediction value. The predicted value ranges from 0 to 1, where 1 indicates that the thermal printer is fully healthy and 0 indicates that it is completely unusable.

10. A computer peripheral device quality assessment system, applied to a computer peripheral device quality assessment method as described in any one of claims 1-9, characterized in that, It includes a processing module, a partitioning module, and an evaluation module; The processing module acquires the time-series data of the thermal printer, uses an adaptive hybrid interpolation strategy to interpolate the missing values ​​of the time-series data to obtain a time-series data matrix, and performs adaptive phased data standardization processing on the time-series data matrix to obtain a standardized feature sequence. The partitioning module performs dynamic selection of health factors on the standardized feature sequences to obtain a health factor dataset, and then performs adaptive data partitioning on the health factor dataset to obtain the partitioned dataset. The evaluation module, based on the first and second datasets, uses an improved AGA-GRU-Informer model to evaluate and predict the health status of thermal printers.

Images