AI-based SMT production line equipment fault prediction method and system

Abstract

This application provides an AI-based SMT production line equipment fault prediction method and system. By collecting and cleaning the operating parameters of multiple devices to construct standardized equipment time-series data, and combining time-domain feature extraction and variational autoencoder reconstruction error modeling, an equipment anomaly score is obtained. Fault warning judgment is performed according to alarm thresholds and the equipment identifier and timestamp are associated and output to the operation and maintenance system. This effectively solves the shortcomings of traditional technologies in multi-device data management, anomaly feature modeling, and fault prediction output, and provides technical support for intelligent fault prediction and operation and maintenance management of SMT production line equipment.

Description

Technical Field

[0001] This application relates to the field of data processing, specifically to an AI-based method and system for predicting equipment failures in SMT production lines. Background Technology

[0002] Existing methods for predicting equipment failures in SMT (Surface Mount Technology) production lines have significant shortcomings. Traditional systems perform poorly in the acquisition and preprocessing of multi-device operating data. They typically lack a unified timestamp alignment and equipment identification management mechanism for the operating parameters of heterogeneous equipment such as pick-and-place machines, reflow ovens, and printers. They also fail to effectively perform missing value filling and outlier marking, resulting in insufficient integrity and consistency of the equipment time-series data, making it difficult to support subsequent refined fault feature extraction.

[0003] Furthermore, existing technologies face bottlenecks in modeling equipment anomaly features. Most systems rely on manually set simple threshold rules to perform fault judgment, lacking the ability to extract time-domain feature matrices through sliding window statistics and feed them into variational autoencoders to perform latent variable distribution modeling. They also cannot utilize reparameterized sampling and encoding / decoding reconstruction mechanisms to learn the potential distribution of normal equipment operation, resulting in insufficient sensitivity and discriminative power of the anomaly score obtained by comparing reconstruction errors with error distribution parameters, thus affecting the accuracy of fault prediction.

[0004] Existing systems have technical shortcomings in the management and output of fault prediction results. They lack a complete closed-loop mechanism for associating and encapsulating fault warnings exceeding alarm thresholds with corresponding device identifiers and timestamps, and then outputting this structured information to the production line maintenance system. This affects the timely response and accurate location of multi-device fault events by maintenance personnel. Solving these problems is crucial for improving the intelligent fault prediction and maintenance management capabilities of SMT production lines. Summary of the Invention

[0005] To address the problems in existing technologies, this application provides an AI-based method and system for predicting SMT production line equipment faults. This method effectively solves the shortcomings of traditional technologies in multi-device data management, anomaly feature modeling, and fault prediction output, providing technical support for intelligent fault prediction and operation and maintenance management of SMT production line equipment.

[0006] To solve at least one of the above problems, this application provides the following technical solution: Firstly, this application provides an AI-based method for predicting equipment failures in SMT production lines, including: The original operating data set is obtained by collecting equipment operating parameters from the surface mount technology production line's pick-and-place machine, reflow oven, and printer, and adding timestamps and equipment identifiers. The original operating data set is then processed by filling missing values ​​and marking outliers according to a preset sampling period to obtain a cleaned data set. The cleaned data set is then grouped by equipment identifier and sorted by timestamp to obtain an equipment time-series data sequence. The device time-series data sequence is processed by performing statistical calculations using a sliding window to obtain a time-domain feature matrix. The time-domain feature matrix is ​​then fed into an encoder network to perform feature encoding and reparameterized sampling to obtain a latent space vector. The latent space vector is then fed into a decoder network to perform reconstruction to obtain reconstructed data. Based on the difference between the reconstructed data and the time-domain feature matrix, a reconstruction error value is calculated. The reconstruction error value is then compared with the error distribution parameters determined during the training phase to obtain an anomaly score. The abnormal score is compared with the preset alarm threshold to obtain the threshold comparison result. Based on the threshold comparison result, the time point that exceeds the preset alarm threshold is marked as the fault warning state to obtain the fault prediction result. The fault prediction result is associated with the corresponding equipment identifier and timestamp, encapsulated, and output to the production line operation and maintenance system.

[0007] Furthermore, it also includes: establishing a data connection with the control system of the pick-and-place machine through an industrial communication protocol and reading the nozzle vacuum degree, placement offset, and feeder feeding status to obtain pick-and-place machine parameter records; establishing a data connection with the control system of the reflow oven through an industrial communication protocol and reading the measured temperature, heating power, and conveyor speed of each temperature zone to obtain reflow oven parameter records; and establishing a data connection with the control system of the printer through an industrial communication protocol and reading the squeegee pressure and printing speed to obtain printer parameter records. The parameter records of the chip mounter, the reflow oven, and the printer are converted to timestamps according to a unified time base to obtain a time-aligned parameter record set. The corresponding equipment identifier and production batch identifier are added to each record in the time-aligned parameter record set to obtain the original operating data set.

[0008] Furthermore, it also includes: performing continuous detection on the original running data set according to a preset sampling period to obtain a missing position index set; performing numerical filling on the missing positions using linear interpolation based on the missing position index set to obtain a filled data set; calculating the local mean and local standard deviation on the filled data set using a sliding window and marking data points that deviate from the local mean by more than a preset multiple of the standard deviation as outliers to obtain a cleaned data set. The cleaned data set is grouped according to the device identifier field to obtain a device group data set. Each group in the device group data set is sorted in ascending order according to the timestamp field to obtain a device time-series data sequence.

[0009] Furthermore, it also includes: dividing the device time-series data sequence into a windowed data segment set by performing windowing according to a preset window length and a preset sliding step size; calculating the mean, standard deviation, maximum and minimum values, skewness and kurtosis of each segment in the windowed data segment set to obtain a window statistics set; and organizing the window statistics set into a matrix according to time order and feature dimensions to obtain a time-domain feature matrix. The temporal feature matrix is ​​fed into the multi-layer convolutional structure of the encoder network to perform layer-by-layer feature extraction to obtain a high-level feature representation. The high-level feature representation is then fed into the distribution mapping layer to perform a linear transformation to obtain the latent variable mean parameter and the latent variable variance parameter. Based on the latent variable mean parameter and the latent variable variance parameter, and by introducing a random noise vector, a reparameterization sampling operation is performed to obtain the latent space vector.

[0010] Furthermore, it also includes: feeding the latent space vector into the multi-layer deconvolution structure of the decoder network to perform layer-by-layer feature expansion to obtain an extended feature representation; feeding the extended feature representation into the output mapping layer to perform dimension restoration to obtain reconstructed data; calculating the point-by-point difference between the reconstructed data and the temporal feature matrix in each feature dimension and taking the absolute value to obtain a point-by-point error matrix; and performing a summation operation on the point-by-point error matrix according to a preset aggregation rule to obtain a reconstruction error value. The mean and standard deviation of the reconstruction error calculated based on normal operating data during the training phase are stored as error distribution parameters to obtain an error distribution parameter record. The difference between the reconstruction error value and the mean reconstruction error in the error distribution parameter record is calculated and divided by the standard deviation of the reconstruction error to obtain a standardized error value. The standardized error value is used as the anomaly score.

[0011] Furthermore, it also includes: comparing the abnormal score with a preset alarm threshold to obtain a single-point comparison result; performing continuous detection on the single-point comparison result according to a time window to obtain a continuous threshold count; and comparing the continuous threshold count with a preset continuous count threshold to obtain a threshold comparison result. Based on the threshold comparison results, the time points when the number of consecutive exceedances reaches a preset consecutive number threshold are marked as fault warning states, and the time points when the number of consecutive exceedances does not reach the preset consecutive number threshold are marked as normal operation states, thus obtaining a state marking sequence. The state marking sequence is then associated and bound with the corresponding timestamps to obtain fault prediction results.

[0012] Furthermore, it also includes: extracting a set of time points marked as fault warning status from the fault prediction results to obtain a set of warning time points; associating and matching the set of warning time points with the corresponding device identifier to obtain a device warning association record; and assembling the device warning association record with the corresponding timestamp and anomaly score value to obtain a warning information data packet. The early warning information data packet is encoded according to a preset data transmission protocol to obtain a standardized transmission message, and the standardized transmission message is sent to the production line operation and maintenance system through a data communication interface.

[0013] Secondly, this application provides an AI-based SMT production line equipment fault prediction system, comprising: The production line data acquisition module is used to collect equipment operating parameters from the chip mounter, reflow oven and printer of the surface mount technology production line, and add timestamps and equipment identifiers to obtain the original operating data set. The original operating data set is filled with missing values ​​and marked with outliers according to a preset sampling period to obtain the cleaned data set. The cleaned data set is grouped by equipment identifier and sorted by timestamp to obtain the equipment time sequence data sequence. The prediction model construction module is used to perform statistical calculations on the device time-series data sequence using a sliding window to obtain a time-domain feature matrix, send the time-domain feature matrix into the encoder network to perform feature encoding and reparameterized sampling to obtain a latent space vector, send the latent space vector into the decoder network to perform reconstruction to obtain reconstructed data, calculate the reconstruction error value based on the difference between the reconstructed data and the time-domain feature matrix, and compare the reconstruction error value with the error distribution parameters determined in the training phase to obtain an anomaly score. The equipment fault prediction module is used to perform a numerical comparison between the anomaly score and a preset alarm threshold to obtain a threshold comparison result, mark the time point that exceeds the preset alarm threshold as a fault warning state based on the threshold comparison result to obtain a fault prediction result, associate and encapsulate the fault prediction result with the corresponding equipment identifier and timestamp, and output it to the production line operation and maintenance system.

[0014] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the AI-based SMT production line equipment fault prediction method.

[0015] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the AI-based SMT production line equipment fault prediction method.

[0016] Fifthly, this application provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the AI-based SMT production line equipment fault prediction method.

[0017] As can be seen from the above technical solution, this application provides an AI-based SMT production line equipment fault prediction method and system. By collecting and cleaning the operating parameters of multiple devices to construct standardized equipment time-series data, and combining time-domain feature extraction and variational autoencoder reconstruction error modeling, the system obtains equipment anomaly scores. The system performs fault warning judgment according to alarm thresholds and outputs the data to the operation and maintenance system along with the equipment identifier and timestamp. This effectively solves the shortcomings of traditional technologies in multi-device data management, anomaly feature modeling, and fault prediction output, and provides technical support for intelligent fault prediction and operation and maintenance management of SMT production line equipment. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the AI-based SMT production line equipment fault prediction method in the embodiments of this application. Figure 2 This is a structural diagram of the AI-based SMT production line equipment fault prediction system in the embodiments of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] The acquisition, storage, use, and processing of data in this application comply with relevant laws and regulations.

[0022] In view of the problems existing in the prior art, this application provides an AI-based SMT production line equipment fault prediction method and system. By collecting and cleaning the operating parameters of multiple devices to construct standardized equipment time-series data, and combining time-domain feature extraction and variational autoencoder reconstruction error modeling, the system obtains equipment anomaly scores. The system performs fault warning judgment according to alarm thresholds and outputs the data to the operation and maintenance system with the associated equipment identifier and timestamp. This effectively solves the shortcomings of traditional technologies in multi-device data management, anomaly feature modeling, and fault prediction output, and provides technical support for intelligent fault prediction and operation and maintenance management of SMT production line equipment.

[0023] To effectively address the shortcomings of traditional technologies in multi-device data management, anomaly feature modeling, and fault prediction output, and to provide technical support for intelligent fault prediction and operation and maintenance management of SMT production line equipment, this application provides an embodiment of an AI-based SMT production line equipment fault prediction method. See [link to embodiment]. Figure 1 The AI-based SMT production line equipment fault prediction method specifically includes the following: Step S101: Collect equipment operating parameters from the surface mount technology production line's pick-and-place machine, reflow oven, and printer, and add timestamps and equipment identifiers to obtain the original operating data set. Perform missing value filling and outlier marking on the original operating data set according to a preset sampling period to obtain the cleaned data set. Group the cleaned data set according to the equipment identifier and sort it according to the timestamp to obtain the equipment time-series data sequence. This embodiment collects equipment operating parameters from the pick-and-place machine, reflow oven, and printer in a surface mount technology production line. The data acquisition process establishes data connections with the control systems of various devices via industrial communication protocols. It reads nozzle vacuum, placement offset, and feeder status from the pick-and-place machine; measured temperatures, heating power, and conveyor speeds for each temperature zone from the reflow oven; and squeegee pressure and printing speed from the printer. While reading each parameter, this embodiment adds a timestamp corresponding to the acquisition time and the device identifier that generated the record to each record, thus aggregating parameter records from different devices to form a raw operating data set.

[0024] Based on the original set of running data, this embodiment performs continuity detection on the data sequence according to a preset sampling period. The detection process uses the timestamp field to determine whether the time interval between adjacent records conforms to the preset sampling period. Locations where the time interval exceeds the preset sampling period are identified as data gaps, and the index position of each gap in the sequence is recorded to form a gap position index set. This gap position index set identifies the data gaps in the original set of running data caused by data acquisition interruptions or communication anomalies.

[0025] Accordingly, this embodiment performs numerical imputation on missing locations based on the missing location index set. The imputation process reads the valid data records adjacent to the missing point, calculates the estimated value of the missing location using linear interpolation, and then fills the estimated value back into the corresponding position. For example, if there is a missing point between two consecutive sampling times in the nozzle vacuum degree sequence of a pick-and-place machine, the arithmetic mean of the two valid vacuum degree values ​​is taken as the imputation value for that missing point. This embodiment obtains the imputed data set after imputing all missing locations in the original running data set.

[0026] Based on the padded dataset, this embodiment performs outlier detection and labeling using a sliding window. During the detection process, the local mean and local standard deviation are calculated for each parameter channel within the sliding window. The deviation of each data point within the window from the local mean is compared with the local standard deviation. Data points deviating from the local mean by more than a preset multiple of the standard deviation are labeled as outliers. Outlier labeling retains the original values ​​using an additional flag field, allowing subsequent analysis to determine whether the anomaly is a data acquisition error or a sudden change in the actual device state. After outlier labeling, this embodiment yields a cleaned dataset.

[0027] Specifically, in this embodiment, the cleaned data set is grouped according to the device identifier field. The grouping process iterates through each record in the cleaned data set, and records from the same device are grouped into the same group based on the device identifier field, forming a device grouped data set. Each group in the device grouped data set corresponds to all the operation records of a specific device on the production line.

[0028] Based on the device grouped data set, this embodiment sorts each group in ascending order according to the timestamp field. The sorting process uses the timestamp value as the sorting basis, organizing the records in each group in order of collection time from earliest to latest, to obtain the device time-series data sequence. The device time-series data sequence maintains the chronological order of the evolution of each device's operating parameters over time, and is used as input for subsequent step S201 to perform sliding window statistical calculations.

[0029] Step S102: Perform statistical calculations on the device time-series data sequence using a sliding window to obtain a time-domain feature matrix. Input the time-domain feature matrix into the encoder network to perform feature encoding and reparameterized sampling to obtain a latent space vector. Input the latent space vector into the decoder network to perform reconstruction to obtain reconstructed data. Calculate the reconstruction error value based on the difference between the reconstructed data and the time-domain feature matrix. Compare the reconstruction error value with the error distribution parameters determined during the training phase to obtain an anomaly score. In this embodiment, the device timing data sequence output in step S101 is divided into windows according to a preset window length and a preset sliding step size. The division process begins at the start of the device timing data sequence, extracting continuous data segments as an analysis window according to the preset window length. Then, the starting position of the window is moved backward by the preset sliding step size, and the next data segment is extracted. This process is repeated until the entire sequence is traversed, resulting in a set of windowed data segments. Each segment in the set of windowed data segments contains records of the device's operating parameters over a continuous time period.

[0030] Based on the aforementioned set of windowed data segments, this embodiment performs statistical calculations on each segment. The calculation process extracts six types of statistics for each parameter channel in each segment: mean, standard deviation, maximum and minimum values, skewness, and kurtosis. The mean reflects the central level of the parameter within the window, the standard deviation reflects the fluctuation range of the parameter, the maximum and minimum values ​​define the range of parameter values, skewness reflects the symmetry of the parameter distribution relative to the mean, and kurtosis reflects the sharpness of the parameter distribution. This embodiment aggregates the statistics of each channel in each segment to form a set of windowed statistics.

[0031] Accordingly, this embodiment organizes the set of window statistics into a matrix according to time order and feature dimension. The organization process uses the time order of the windows as the row index of the matrix and the various statistics of each parameter channel as the column index, filling the corresponding statistical values ​​into the appropriate positions in the matrix to obtain the temporal feature matrix. This temporal feature matrix compresses the original high-frequency sampling data into a structured feature representation, reducing the computational burden on the subsequent encoder network.

[0032] Based on the temporal feature matrix, this embodiment feeds it into the multi-layer convolutional structure of the encoder network to perform layer-by-layer feature extraction. Each convolutional layer of the encoder network captures local correlation patterns in the feature dimension through a sliding convolutional kernel, compressing the feature dimension and expanding the channel dimension layer by layer to output a high-level feature representation. In this embodiment, the high-level feature representation is fed into a distribution mapping layer to perform a linear transformation. The distribution mapping layer contains two parallel projection branches, which output the latent variable mean parameter and the latent variable variance parameter, respectively.

[0033] Specifically, this embodiment performs reparameterization sampling based on the latent variable mean and variance parameters. The sampling process generates a random noise vector with the same dimension as the latent variables. This random noise vector is multiplied by the square root of the latent variable variance parameter and then added to the latent variable mean parameter to obtain the latent space vector. The reparameterization technique transfers the randomness of the sampling to external noise input, making the gradient of the latent space vector relative to the latent variable mean and variance parameters computable.

[0034] Based on the latent space vector, this embodiment feeds it into a multi-layer deconvolutional structure of the decoder network to perform layer-by-layer feature expansion. Each deconvolutional layer of the decoder network upsamples and expands the feature representation in terms of dimension, restoring it layer by layer to the same dimensional structure as the temporal feature matrix, and outputs the expanded feature representation. This embodiment feeds the expanded feature representation into the output mapping layer to perform dimensionality restoration, obtaining the reconstructed data. The reconstructed data has the same dimensional structure as the temporal feature matrix.

[0035] Accordingly, this embodiment calculates the pointwise difference between the reconstructed data and the temporal feature matrix at each feature dimension and takes the absolute value to obtain the pointwise error matrix. This embodiment performs a summation operation on the pointwise error matrix according to a preset aggregation rule, and accumulates the absolute errors of all elements in the matrix to obtain the reconstruction error value. The reconstruction error value reflects the reconstruction accuracy of the decoder network for the current input sample.

[0036] Based on the reconstructed error value, this embodiment compares it with the error distribution parameters determined during the training phase. During the training phase, the mean and standard deviation of the reconstruction error are calculated based on normal operating data, and both are recorded and stored as error distribution parameters. This embodiment calculates the difference between the current reconstruction error value and the mean reconstruction error in the recorded error distribution parameters, divides the difference by the standard deviation of the reconstruction error to obtain a standardized error value, and uses this standardized error value as an anomaly score. This anomaly score is used as input for the subsequent threshold comparison in step S103.

[0037] Step S103: Perform a numerical comparison between the anomaly score and the preset alarm threshold to obtain the threshold comparison result. Based on the threshold comparison result, mark the time point that exceeds the preset alarm threshold as a fault warning state to obtain the fault prediction result. Associate and encapsulate the fault prediction result with the corresponding equipment identifier and timestamp and output it to the production line operation and maintenance system.

[0038] This embodiment compares the anomaly score output in step S102 with a preset alarm threshold. The comparison process determines whether the anomaly score for each time window exceeds the preset alarm threshold, recording the comparison result as either "exceeds" or "does not exceed," thus obtaining a single-point comparison result. The preset alarm threshold is configured based on the anomaly score distribution of the normal operating data during the training phase and the acceptable false alarm rate level for the business.

[0039] Based on the single-point comparison results, this embodiment performs continuous detection according to time windows. The detection process iterates through the single-point comparison results at each time point along the time axis, and counts the number of time points where the state continuously exceeds the preset alarm threshold, thus obtaining a statistical quantity of consecutive threshold exceedances. For example, if a pick-and-place machine's abnormal scores exceed the preset alarm threshold for five consecutive time windows, then the statistical quantity of consecutive threshold exceedances for that period is five. Continuous detection can filter out isolated threshold exceedances caused by instantaneous fluctuations, reducing the false alarm rate.

[0040] Accordingly, this embodiment compares the consecutive threshold exceedance statistics with a preset consecutive exceedance threshold to obtain a threshold comparison result. The comparison process determines whether the consecutive threshold exceedance statistics reach the preset consecutive exceedance threshold; if so, it is determined that an alarm condition has been triggered; otherwise, it is determined that an alarm condition has not been triggered. The preset consecutive exceedance threshold is configured based on the typical evolution duration of the equipment fault and the alarm response timeliness requirements.

[0041] Based on the threshold comparison results, this embodiment performs fault warning status marking. The marking process marks the time point when the number of consecutive exceedances reaches a preset consecutive number threshold as a fault warning state, and the time point when the number of consecutive exceedances does not reach the preset consecutive number threshold as a normal operation state, resulting in a status marking sequence. This embodiment associates and binds the status marking sequence with corresponding timestamps, establishing a correspondence between each status mark and a specific time point, thus obtaining the fault prediction result.

[0042] Specifically, in this embodiment, a set of warning time points is obtained by extracting the set of time points marked as fault warning states from the fault prediction results. The extraction process iterates through each record in the fault prediction results, filters records whose status marker field is a fault warning state, and aggregates their corresponding time point information. This embodiment then associates and matches the set of warning time points with the corresponding device identifiers to establish a correspondence between the warning information and the specific device that generated the warning, thus obtaining device warning association records.

[0043] Based on the aforementioned device early warning association records, this embodiment assembles them with corresponding timestamps and anomaly score values. The assembly process encapsulates the device identifier, early warning time point, and anomaly score value into a structured data record, resulting in an early warning information data packet. This early warning information data packet contains complete information needed by maintenance personnel to locate faulty equipment and determine the severity of the fault.

[0044] Accordingly, in this embodiment, the early warning information data packet is encoded into a standardized transmission message according to a preset data transmission protocol. The encoding process converts the early warning information data packet into a message structure that conforms to the protocol requirements, based on the interface specifications of the production line operation and maintenance system. In this embodiment, the standardized transmission message is sent to the production line operation and maintenance system through a data communication interface, allowing maintenance personnel to perform equipment verification and maintenance scheduling based on the fault early warning information.

[0045] As can be seen from the above description, the AI-based SMT production line equipment fault prediction method provided in this application can construct standardized equipment time-series data by collecting and cleaning multiple equipment operating parameters, and obtain equipment anomaly scores by combining time-domain feature extraction and variational autoencoder reconstruction error modeling. It can also perform fault warning judgment according to alarm thresholds and output the data to the operation and maintenance system by associating equipment identifiers and timestamps. This effectively solves the shortcomings of traditional technologies in multi-equipment data management, anomaly feature modeling, and fault prediction output, and provides technical support for intelligent fault prediction and operation and maintenance management of SMT production line equipment.

[0046] In one embodiment of the AI-based SMT production line equipment fault prediction method of this application, it may further include the following: Step S201: Establish a data connection with the control system of the pick-and-place machine through the industrial communication protocol and read the nozzle vacuum degree, placement offset and feeder feeding status to obtain the pick-and-place machine parameter record; establish a data connection with the control system of the reflow oven through the industrial communication protocol and read the measured temperature, heating power and conveyor speed of each temperature zone to obtain the reflow oven parameter record; establish a data connection with the control system of the printer through the industrial communication protocol and read the squeegee pressure and printing speed to obtain the printer parameter record. Step S202: Perform timestamp conversion on the chip mounter parameter record, the reflow oven parameter record, and the printer parameter record according to a unified time base to obtain a time-aligned parameter record set. Add corresponding equipment identifiers and production batch identifiers to each record in the time-aligned parameter record set to obtain the original operating data set.

[0047] This embodiment establishes a data connection with the pick-and-place machine's control system via an industrial communication protocol. The connection process configures protocol parameters and access addresses based on the open communication interface of the pick-and-place machine's control system, establishing a stable data transmission channel. This embodiment reads three types of parameters from the pick-and-place machine's control system via this data connection: nozzle vacuum level, placement offset, and feeder feeding status. Nozzle vacuum level reflects the nozzle's sealing performance and pickup capability; placement offset reflects the degree of deviation of the component placement position from the target coordinates; and feeder feeding status reflects the working status of the feeding mechanism. This embodiment records and organizes the three types of parameters according to the acquisition time to obtain the pick-and-place machine parameter record.

[0048] After the pick-and-place machine parameters are recorded and collected, this embodiment establishes a data connection with the reflow oven control system via an industrial communication protocol. The connection process uses the same protocol configuration as the pick-and-place machine, adapted to the interface characteristics of the reflow oven control system. This embodiment reads three types of parameters from the reflow oven control system via this data connection: measured temperature of each temperature zone, heating power, and conveyor speed. The measured temperature of each temperature zone reflects the actual temperature levels of the preheating zone, constant temperature zone, and reflow zone within the oven chamber; the heating power reflects the output status of the heating elements in each temperature zone; and the conveyor speed reflects the speed at which the printed circuit board is transported within the oven chamber. This embodiment records and organizes the three types of parameters according to the time of acquisition, obtaining the reflow oven parameter record.

[0049] Accordingly, this embodiment establishes a data connection with the printing press control system via an industrial communication protocol. The connection process completes protocol handshake and channel establishment according to the interface specifications of the printing press control system. This embodiment reads two types of parameters from the printing press control system via this data connection: squeegee pressure and printing speed. Squeegee pressure reflects the pressure level applied by the squeegee to the stencil during the printing stroke, while printing speed reflects the speed at which the squeegee completes a single printing stroke. This embodiment records and organizes the two types of parameters according to the acquisition time to obtain the printing press parameter record.

[0050] After the parameter records of the pick-and-place machine, reflow oven, and printer are collected, this embodiment performs timestamp conversion on the three types of parameter records according to a unified time base. The conversion process first determines a unified time base at the production line level, and then converts the original timestamps in each equipment parameter record to a standard timestamp under this unified time base. For example, if there is a clock deviation between the control systems of the pick-and-place machine and the reflow oven, their respective timestamps are corrected according to a pre-calibrated clock deviation amount, so that parameter records collected at the same physical moment have the same standard timestamp. After completing the timestamp conversion, this embodiment obtains a time-aligned parameter record set.

[0051] Specifically, in this embodiment, a corresponding device identifier and production batch identifier are appended to each record in the time alignment parameter record set. The device identifier indicates which specific machine on the production line the record originates from, and the production batch identifier indicates the production task batch corresponding to the record. The appending process iterates through each record in the time alignment parameter record set, filling the device identifier field and the production batch identifier field according to the record's source device and the production batch at which the data was collected. After the identifier appending is completed, this embodiment obtains the original running data set, which is used as input for subsequent steps S301 to perform missing value filling and outlier marking.

[0052] In one embodiment of the AI-based SMT production line equipment fault prediction method of this application, it may further include the following: Step S301: Perform continuity detection on the original running data set according to a preset sampling period to obtain a missing position index set. Based on the missing position index set, perform numerical filling on the missing positions using linear interpolation to obtain a filled data set. Calculate the local mean and local standard deviation on the filled data set using a sliding window and mark data points that deviate from the local mean by more than a preset multiple of the standard deviation as outliers to obtain a cleaned data set. Step S302: Perform a grouping operation on the cleaned data set according to the device identifier field to obtain a device group data set, and sort each group in the device group data set in ascending order according to the timestamp field to obtain a device time-series data sequence.

[0053] In this embodiment, the original running data set output in step S202 is subjected to continuity detection according to a preset sampling period. The detection process calculates the time interval between adjacent records based on the timestamp sequence of each parameter channel and compares this time interval with the preset sampling period. When the time interval between adjacent records exceeds the preset sampling period, it is determined that there is a data gap at that location, and the index of the missing location in the sequence is registered in the missing location index set. The missing location index set records all data gaps in the original running data set caused by communication interruptions or abnormal acquisition.

[0054] Based on the missing location index set, this embodiment uses linear interpolation to fill in the missing locations. The filling process iterates through each index position in the missing location index set, reads the valid data records immediately before and after that position, and calculates an estimated value for the missing location based on the values ​​of the preceding and following valid data points and the time interval. Linear interpolation assumes that the parameters change linearly within the missing interval, and uses the value at the corresponding time on the line connecting the preceding and following data points as the fill value. This embodiment fills all missing locations in the original running dataset to obtain the filled dataset.

[0055] Accordingly, this embodiment performs outlier detection on the filled dataset using a sliding window. The detection process involves setting a sliding window on each parameter channel, with the window moving gradually along the time axis. At each window position, the local mean and local standard deviation of the data points within the window are calculated. This embodiment divides the absolute value of the difference between each data point within the window and the local mean by the local standard deviation to obtain the standardized deviation of that data point. For example, if the standardized deviation of the measured temperature in a reflow oven temperature zone exceeds a preset multiple at a certain moment, then that data point is identified as a suspected outlier.

[0056] After the standardized deviation calculation is completed, this embodiment marks data points that deviate from the local mean by more than a preset multiple of the standard deviation as outliers. The marking process involves adding an anomaly flag field to data points that meet the anomaly criteria, while retaining the original values ​​without modification. The anomaly flag field is used for subsequent analysis to determine whether the anomaly originates from a data acquisition error or a sudden change in the actual device state. After completing the outlier marking, this embodiment obtains a cleaned dataset, which contains a complete parameter sequence processed by missing value imputation and anomaly marking.

[0057] Specifically, in this embodiment, the cleaned data set is grouped according to the device identifier field. The grouping process iterates through each record in the cleaned data set, reads the device identifier field value of each record, and groups records with the same device identifier into the same group. After grouping, this embodiment obtains a device grouped data set, where each group corresponds to all operating parameter records of a specific device in the surface mount technology production line.

[0058] Based on the device grouped data set, this embodiment sorts each group in ascending order according to the timestamp field. The sorting process uses the value of the timestamp field as the sorting basis, reorganizing the records in each group in order of collection time from earliest to latest. After the sorting is completed, this embodiment obtains a device time-series data sequence. The device time-series data sequence maintains the chronological relationship of the evolution of the operating parameters of each device over time, and is used as input for subsequent step S401 to perform sliding window statistical calculations.

[0059] In one embodiment of the AI-based SMT production line equipment fault prediction method of this application, it may further include the following: Step S401: Perform windowing on the device time-series data sequence according to the preset window length and preset sliding step size to obtain a set of windowed data segments. Calculate the mean, standard deviation, maximum and minimum values, skewness and kurtosis of each segment in the set of windowed data segments to obtain a set of window statistics. Organize the set of window statistics into a matrix according to time order and feature dimensions to obtain a time-domain feature matrix. Step S402: The temporal feature matrix is ​​fed into the multi-layer convolutional structure of the encoder network to perform layer-by-layer feature extraction to obtain a high-level feature representation. The high-level feature representation is fed into the distribution mapping layer to perform a linear transformation to obtain the latent variable mean parameter and the latent variable variance parameter. Based on the latent variable mean parameter and the latent variable variance parameter, and by introducing a random noise vector, a reparameterization sampling operation is performed to obtain the latent space vector.

[0060] In this embodiment, the device timing data sequence output in step S302 is divided into windows according to a preset window length and a preset sliding step size. The division process starts from the beginning of the device timing data sequence, and a continuous data record is extracted as the first window according to the preset window length. Then, the starting position of the window is moved backward according to the preset sliding step size, and the next data record is extracted as the second window. This embodiment repeats the window extraction and position movement operation until the end position of the window reaches the end of the sequence, and all extracted data segments are gathered to form a windowed data segment set.

[0061] Based on the aforementioned set of windowed data segments, this embodiment performs statistical calculations on each segment. The calculation process extracts six types of statistics for each parameter channel within each segment. The mean is obtained by calculating the arithmetic mean of all data points within the window, reflecting the central level of the parameter during that time period. The standard deviation is obtained by calculating the root mean of the sum of the squares of the differences between each data point and the mean, reflecting the dispersion and fluctuation range of the parameter. The maximum and minimum values ​​are taken as the upper and lower bounds of the data points within the window, respectively, defining the range of parameter values.

[0062] Accordingly, this embodiment further calculates the skewness and kurtosis statistics for each parameter channel of each segment. Skewness is obtained by calculating and standardizing the third central moment of the difference between each data point and the mean, reflecting the symmetry of the parameter distribution relative to the mean. Positive skewness indicates a longer right tail, and negative skewness indicates a longer left tail. Kurtosis is obtained by calculating and standardizing the fourth central moment of the difference between each data point and the mean, reflecting the sharpness and tail thickness of the parameter distribution. This embodiment aggregates the six types of statistics for each channel of each segment into a window statistics set.

[0063] Based on the aforementioned set of window statistics, this embodiment organizes them into a matrix according to chronological order and feature dimensions. The organization process uses the chronological order of the windows as the row indices of the matrix and the various statistics of each parameter channel as the column indices, filling the corresponding statistical values ​​into the appropriate positions in the matrix. For example, if the device time-series data sequence contains three parameter channels and six types of statistics are extracted from each channel, then the matrix has eighteen columns. After completing the matrix organization, this embodiment obtains a time-domain feature matrix.

[0064] Specifically, in this embodiment, the temporal feature matrix is ​​fed into a multi-layer convolutional structure of the encoder network to perform layer-by-layer feature extraction. The first convolutional layer of the encoder network receives the temporal feature matrix as input and captures local correlation patterns along the feature dimension through a sliding convolutional kernel. Each convolutional layer is followed by an activation function layer to introduce non-linear transformation capabilities, compressing the feature dimension and expanding the channel dimension layer by layer. In this embodiment, the temporal feature matrix is ​​processed by each convolutional layer of the encoder network to obtain a high-level feature representation, which encodes the original statistical features into a compact abstract representation.

[0065] Based on the high-level feature representation, this embodiment feeds it into a distribution mapping layer for linear transformation. The distribution mapping layer contains two parallel fully connected projection branches. The first branch performs linear projection on the high-level feature representation to obtain the latent variable mean parameter, and the second branch performs linear projection on the high-level feature representation to obtain the latent variable variance parameter. This embodiment performs reparameterized sampling based on the latent variable mean parameter and the latent variable variance parameter. The sampling process generates a random noise vector with the same dimension as the latent variables and following a standard normal distribution. The random noise vector is multiplied by the square root of the latent variable variance parameter and then added to the latent variable mean parameter to obtain the latent space vector. This latent space vector is used as input for subsequent decoding and reconstruction in step S501.

[0066] In one embodiment of the AI-based SMT production line equipment fault prediction method of this application, it may further include the following: Step S501: The latent space vector is fed into the multi-layer deconvolution structure of the decoder network to perform layer-by-layer feature expansion to obtain the extended feature representation. The extended feature representation is fed into the output mapping layer to perform dimension restoration to obtain the reconstructed data. The point-by-point difference between the reconstructed data and the temporal feature matrix in each feature dimension is calculated and the absolute value is taken to obtain the point-by-point error matrix. The point-by-point error matrix is ​​summed according to the preset aggregation rule to obtain the reconstruction error value. Step S502: Store the mean reconstruction error and standard deviation of reconstruction error calculated based on normal operation data during the training phase as error distribution parameters to obtain an error distribution parameter record. Perform difference calculation between the reconstruction error value and the mean reconstruction error in the error distribution parameter record and divide by the standard deviation of reconstruction error to obtain a standardized error value. Use the standardized error value as an anomaly score.

[0067] In this embodiment, the latent space vector output in step S402 is fed into the multi-layer deconvolution structure of the decoder network to perform layer-by-layer feature expansion. The decoder network adopts a hierarchical structure symmetrical to the encoder network, with each layer consisting of deconvolution operation units and activation function units. After entering the decoder network, the latent space vector is processed through each deconvolution layer in sequence. Each deconvolution operation upsamples and expands the feature representation in terms of dimension, restoring the original dimensional structure of the data layer by layer to obtain the expanded feature representation.

[0068] Based on the extended feature representation, this embodiment feeds it into the output mapping layer for dimensionality reduction. The output mapping layer compresses the channel dimensions of the extended feature representation to the same number of feature dimensions as the temporal feature matrix through a linear transformation, outputting a tensor with a dimensional structure completely consistent with the temporal feature matrix, thus obtaining the reconstructed data. The reconstructed data represents the reconstruction result of the decoder network after latent space encoding of the input samples.

[0069] Accordingly, this embodiment calculates the pointwise difference between the reconstructed data and the temporal feature matrix output in step S401 at each feature dimension. The calculation process involves subtracting the values ​​of the reconstructed data and the temporal feature matrix at corresponding positions, taking the absolute value of the difference to eliminate the influence of positive and negative directions, and obtaining a pointwise error matrix. Each element in the pointwise error matrix reflects the degree of deviation between the reconstructed data and the original input at that position.

[0070] Based on the pointwise error matrix, this embodiment performs a summation operation according to a preset aggregation rule. The summation process traverses all elements in the pointwise error matrix, summing the absolute error values ​​of each element to obtain the reconstruction error value. This reconstruction error value compresses the multidimensional error in matrix form into a scalar form, reflecting the overall reconstruction accuracy of the decoder network for the current input sample. Device data under normal operating conditions conforms to the data distribution characteristics learned during the training phase, allowing the decoder network to accurately reconstruct the data with a small reconstruction error value. Device data under abnormal conditions deviates from the normal distribution characteristics, making accurate reconstruction difficult for the decoder network and increasing the reconstruction error value.

[0071] Specifically, in this embodiment, the statistical distribution parameters of the reconstruction error are calculated based on normal operating data during the training phase. The training process uses historical data samples containing only the normal operating state of the device to train the encoder and decoder networks. After training, all normal operating samples are fed into the trained network to perform forward inference, collecting the reconstruction error values ​​for each sample. This embodiment calculates the arithmetic mean of the collected reconstruction error values ​​to obtain the reconstruction error mean, calculates the standard deviation to obtain the reconstruction error standard deviation, and stores both as error distribution parameters to obtain the error distribution parameter record.

[0072] Based on the recorded error distribution parameters, this embodiment calculates the difference between the current reconstruction error value and the mean reconstruction error in the recorded error distribution parameters. The difference calculation yields the deviation of the current reconstruction error from the normal error center level. This embodiment divides the deviation by the standard deviation of the reconstruction error to perform standardization, obtaining a standardized error value. Standardization eliminates scale fluctuations in the reconstruction error caused by batch differences in data, making the error indicators of different samples comparable. This embodiment uses the standardized error value as an anomaly score, which is used as input for the threshold comparison in subsequent step S601.

[0073] In one embodiment of the AI-based SMT production line equipment fault prediction method of this application, it may further include the following: Step S601: Compare the abnormal score with the preset alarm threshold to obtain a single-point comparison result. Perform continuous detection on the single-point comparison result according to the time window to obtain a continuous threshold count. Compare the continuous threshold count with the preset continuous count threshold to obtain a threshold comparison result. Step S602: Based on the threshold comparison results, mark the time points when the number of consecutive exceedances reaches the preset consecutive number threshold as fault warning state and mark the time points when the number of consecutive exceedances does not reach the preset consecutive number threshold as normal operation state to obtain a state marking sequence. Associate and bind the state marking sequence with the corresponding timestamp to obtain the fault prediction result.

[0074] This embodiment compares the anomaly score output in step S502 with a preset alarm threshold. The comparison process determines whether the anomaly score for each time window exceeds the preset alarm threshold, recording the result as either "exceeds" or "does not exceed," thus obtaining a single-point comparison result. The preset alarm threshold is configured based on the anomaly score distribution characteristics of the normal operating data during the training phase, and the configuration must balance the relationship between the false positive rate and the false negative rate.

[0075] Based on the single-point comparison results, this embodiment performs continuous detection according to a time window. The detection process sequentially traverses the single-point comparison results at each time point along the time axis. When a time point with a state of "exceeding" is encountered, counting begins, and the number of consecutive time points with the "exceeding" state is counted. When a time point with a state of "not exceeding" is encountered, the counter is reset to zero and the counting restarts. This embodiment records the number of consecutive time points with the "exceeding" state within each time period as a statistical measure of consecutive threshold exceedances.

[0076] Accordingly, this embodiment compares the consecutive threshold exceedance statistics with a preset consecutive exceedance threshold to obtain a threshold comparison result. The comparison process determines whether the consecutive threshold exceedance statistics reach the preset consecutive exceedance threshold. If they do, the alarm triggering condition is met; otherwise, it is not. For example, if the preset consecutive exceedance threshold is configured to three times, an alarm will only be triggered when the anomaly score of a device exceeds the preset alarm threshold for three or more consecutive time windows. The continuity condition can filter out isolated threshold exceedances caused by instantaneous disturbances, reducing false alarms and interference with maintenance personnel.

[0077] Based on the threshold comparison results, this embodiment performs fault warning status marking. The marking process iterates through the threshold comparison results at each time point, marking the time point where the number of consecutive threshold exceedances reaches a preset consecutive number threshold as a fault warning state, and marking the time point where the number of consecutive threshold exceedances does not reach the preset consecutive number threshold as a normal operating state. In this embodiment, the status markings of all time points are organized in chronological order to form a status marking sequence, and the status marking sequence records the operating status determination results of the device at each time point.

[0078] Specifically, in this embodiment, the status marker sequence is associated and bound with the corresponding timestamps. The binding process establishes a one-to-one correspondence between each status marker in the status marker sequence and the timestamp corresponding to the time window in which the marker was generated, thus associating each fault warning state or normal operating state with specific time information. After completing the association and binding, this embodiment obtains a fault prediction result, which includes the status markers and corresponding timestamp information at each time point, for use as input in subsequent steps S701 when performing warning information extraction and encapsulation.

[0079] In one embodiment of the AI-based SMT production line equipment fault prediction method of this application, it may further include the following: Step S701: Extract the set of time points marked as fault warning status from the fault prediction results to obtain the warning time point set; associate and match the warning time point set with the corresponding device identifier to obtain the device warning association record; assemble the device warning association record with the corresponding timestamp and abnormal score value to obtain the warning information data packet. Step S702: Encode the early warning information data packet according to a preset data transmission protocol to obtain a standardized transmission message, and send the standardized transmission message to the production line operation and maintenance system through the data communication interface.

[0080] This embodiment extracts the set of time points marked as fault warning states from the fault prediction results output in step S602. The extraction process iterates through each record in the fault prediction results, reads the status marker field of each record, filters records whose status marker field value is a fault warning state, and aggregates their corresponding time point information to obtain the warning time point set. The warning time point set contains all time points determined to have a fault risk.

[0081] Based on the set of warning time points, this embodiment associates and matches them with the corresponding equipment identifiers. The matching process traces the equipment identifier information corresponding to each time point in the original data according to the source records of each time point in the warning time point set, establishing a correspondence between the warning time point and the specific equipment that generated the warning. After completing the association matching, this embodiment obtains an equipment warning association record, which clearly identifies which piece of equipment on the production line triggered each fault warning.

[0082] Accordingly, this embodiment assembles the equipment warning association records with their corresponding timestamps and anomaly score values. The assembly process involves reading the precise timestamp and anomaly score value for each record in the equipment warning association records, and encapsulating the equipment identifier, timestamp, and anomaly score value into a structured data record. For example, if a pick-and-place machine triggers a fault warning at a certain time and the corresponding anomaly score value is high, the assembled record will contain three fields: the pick-and-place machine's equipment identifier, the timestamp at that time, and the anomaly score value. After completing the field assembly, this embodiment obtains a warning information data packet.

[0083] Based on the aforementioned early warning information data packet, this embodiment encodes it according to a preset data transmission protocol. The encoding process serializes each field in the early warning information data packet according to the protocol-defined format and order, based on the interface specifications of the production line operation and maintenance system. The required message header and verification information are added to obtain a standardized transmission message. This standardized transmission message conforms to the data reception specifications of the production line operation and maintenance system and can be correctly parsed and processed by the system.

[0084] Specifically, in this embodiment, the standardized transmission message is sent to the production line operation and maintenance system via a data communication interface. The sending process invokes a pre-configured data communication interface, writes the standardized transmission message into the communication channel, and transmits it to the data receiving port of the production line operation and maintenance system. After receiving and parsing the standardized transmission message, the production line operation and maintenance system extracts the equipment identifier, timestamp, and anomaly score information, allowing maintenance personnel to locate abnormal equipment and arrange verification and maintenance work based on fault warning information.

[0085] To effectively address the shortcomings of traditional technologies in multi-device data management, anomaly feature modeling, and fault prediction output, and to provide technical support for intelligent fault prediction and operation and maintenance management of SMT production line equipment, this application provides an embodiment of an AI-based SMT production line equipment fault prediction system for implementing all or part of the aforementioned AI-based SMT production line equipment fault prediction method. See [link to embodiment]. Figure 2 The AI-based SMT production line equipment fault prediction system specifically includes the following components: The production line data acquisition module 10 is used to collect equipment operating parameters from the chip mounter, reflow oven and printer of the surface mount technology production line and add timestamps and equipment identifiers to obtain the original operating data set. The original operating data set is filled with missing values ​​and marked with outliers according to a preset sampling period to obtain the cleaned data set. The cleaned data set is grouped according to the equipment identifier and sorted according to the timestamp to obtain the equipment time sequence data sequence. The prediction model construction module 20 is used to perform statistical calculations on the device time series data sequence using a sliding window to obtain a time-domain feature matrix, send the time-domain feature matrix into the encoder network to perform feature encoding and reparameterized sampling to obtain a latent space vector, send the latent space vector into the decoder network to perform reconstruction to obtain reconstructed data, calculate the reconstruction error value based on the difference between the reconstructed data and the time-domain feature matrix, and compare the reconstruction error value with the error distribution parameters determined in the training phase to obtain an anomaly score. The equipment fault prediction module 30 is used to perform a numerical comparison between the anomaly score and the preset alarm threshold to obtain the threshold comparison result, mark the time point that exceeds the preset alarm threshold as a fault warning state based on the threshold comparison result to obtain the fault prediction result, associate and encapsulate the fault prediction result with the corresponding equipment identifier and timestamp, and output it to the production line operation and maintenance system.

[0086] As can be seen from the above description, the AI-based SMT production line equipment fault prediction system provided in this application embodiment can construct standardized equipment time-series data through the collection of multi-equipment operating parameters and time-series cleaning, obtain equipment anomaly scores by combining time-domain feature extraction and variational autoencoder reconstruction error modeling, perform fault warning judgment according to alarm thresholds, and output the system with equipment identifier and timestamp. This effectively solves the shortcomings of traditional technologies in multi-equipment data management, anomaly feature modeling, and fault prediction output, and provides technical support for intelligent fault prediction and operation and maintenance management of SMT production line equipment.

[0087] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the AI-based SMT production line equipment fault prediction method.

[0088] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described AI-based SMT production line equipment fault prediction method.

[0089] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described AI-based SMT production line equipment fault prediction method.

[0090] In this embodiment of the invention, standardized equipment time-series data is constructed by collecting and cleaning operating parameters of multiple devices. Equipment anomaly scores are obtained by combining time-domain feature extraction and variational autoencoder reconstruction error modeling. Fault warning judgment is performed according to alarm thresholds and output to the operation and maintenance system with associated equipment identifiers and timestamps. This effectively solves the shortcomings of traditional technologies in multi-device data management, anomaly feature modeling, and fault prediction output, and provides technical support for intelligent fault prediction and operation and maintenance management of SMT production line equipment.

[0091] 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 embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0092] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0093] 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.

[0094] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0095] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A fault prediction method for SMT production line equipment based on AI, characterized in that, The method includes: The original operating data set is obtained by collecting equipment operating parameters from the surface mount technology production line's pick-and-place machine, reflow oven, and printer, and adding timestamps and equipment identifiers. The original operating data set is then processed by filling missing values ​​and marking outliers according to a preset sampling period to obtain a cleaned data set. The cleaned data set is then grouped by equipment identifier and sorted by timestamp to obtain an equipment time-series data sequence. The device time-series data sequence is processed by performing statistical calculations using a sliding window to obtain a time-domain feature matrix. The time-domain feature matrix is ​​then fed into an encoder network to perform feature encoding and reparameterized sampling to obtain a latent space vector. The latent space vector is then fed into a decoder network to perform reconstruction to obtain reconstructed data. Based on the difference between the reconstructed data and the time-domain feature matrix, a reconstruction error value is calculated. The reconstruction error value is then compared with the error distribution parameters determined during the training phase to obtain an anomaly score. The abnormal score is compared with the preset alarm threshold to obtain the threshold comparison result. Based on the threshold comparison result, the time point that exceeds the preset alarm threshold is marked as the fault warning state to obtain the fault prediction result. The fault prediction result is associated with the corresponding equipment identifier and timestamp, encapsulated, and output to the production line operation and maintenance system.

2. The AI-based SMT production line equipment fault prediction method according to claim 1, characterized in that, The raw operating data set is obtained by collecting equipment operating parameters from the surface mount technology production line's pick-and-place machines, reflow ovens, and printers, and adding timestamps and equipment identifiers. This includes: The system establishes a data connection with the control system of the pick-and-place machine through the industrial communication protocol and reads the nozzle vacuum degree, placement offset and feeder feeding status to obtain the pick-and-place machine parameter record. The system establishes a data connection with the control system of the reflow oven through the industrial communication protocol and reads the measured temperature, heating power and conveyor speed of each temperature zone to obtain the reflow oven parameter record. The system establishes a data connection with the control system of the printer through the industrial communication protocol and reads the squeegee pressure and printing speed to obtain the printer parameter record. The parameter records of the chip mounter, the reflow oven, and the printer are converted to timestamps according to a unified time base to obtain a time-aligned parameter record set. The corresponding equipment identifier and production batch identifier are added to each record in the time-aligned parameter record set to obtain the original operating data set.

3. The AI-based SMT production line equipment fault prediction method according to claim 1, characterized in that, The process involves filling in missing values ​​and marking outliers in the original operational data set according to a preset sampling period to obtain a cleaned data set. The cleaned data set is then grouped by device identifier and sorted by timestamp to obtain a device time-series data sequence, including: A set of missing position indices is obtained by performing continuous detection on the original running data set according to a preset sampling period. Based on the missing position indices, numerical filling is performed on the missing positions using linear interpolation to obtain a filled data set. The local mean and local standard deviation are calculated on the filled data set using a sliding window, and data points that deviate from the local mean by more than a preset multiple of the standard deviation are marked as outliers to obtain a cleaned data set. The cleaned data set is grouped according to the device identifier field to obtain a device group data set. Each group in the device group data set is sorted in ascending order according to the timestamp field to obtain a device time-series data sequence.

4. The AI-based SMT production line equipment fault prediction method according to claim 1, characterized in that, The step of performing statistical calculations on the device's time-series data sequence using a sliding window to obtain a temporal feature matrix, and then feeding the temporal feature matrix into the encoder network to perform feature encoding and reparameterized sampling to obtain a latent space vector, includes: The device time-series data sequence is divided into windowed data segments according to a preset window length and a preset sliding step size. The mean, standard deviation, maximum and minimum values, skewness and kurtosis of each segment in the windowed data segment set are calculated to obtain a set of window statistics. The set of window statistics is organized into a matrix according to time order and feature dimension to obtain a time-domain feature matrix. The temporal feature matrix is ​​fed into the multi-layer convolutional structure of the encoder network to perform layer-by-layer feature extraction to obtain a high-level feature representation. The high-level feature representation is then fed into the distribution mapping layer to perform a linear transformation to obtain the latent variable mean parameter and the latent variable variance parameter. Based on the latent variable mean parameter and the latent variable variance parameter, and by introducing a random noise vector, a reparameterization sampling operation is performed to obtain the latent space vector.

5. The AI-based SMT production line equipment fault prediction method according to claim 1, characterized in that, The process involves feeding the latent space vector into a decoder network to perform reconstruction, obtaining reconstructed data, calculating a reconstruction error value based on the difference between the reconstructed data and the temporal feature matrix, and comparing the reconstruction error value with the error distribution parameters determined during the training phase to obtain an anomaly score. This includes: The latent space vector is fed into the multi-layer deconvolution structure of the decoder network to perform layer-by-layer feature expansion to obtain the extended feature representation. The extended feature representation is fed into the output mapping layer to perform dimension restoration to obtain the reconstructed data. The point-by-point difference between the reconstructed data and the temporal feature matrix in each feature dimension is calculated and the absolute value is taken to obtain the point-by-point error matrix. The point-by-point error matrix is ​​summed according to the preset aggregation rule to obtain the reconstruction error value. The mean and standard deviation of the reconstruction error calculated based on normal operating data during the training phase are stored as error distribution parameters to obtain an error distribution parameter record. The difference between the reconstruction error value and the mean reconstruction error in the error distribution parameter record is calculated and divided by the standard deviation of the reconstruction error to obtain a standardized error value. The standardized error value is used as the anomaly score.

6. The AI-based SMT production line equipment fault prediction method according to claim 1, characterized in that, The step of comparing the anomaly score with a preset alarm threshold to obtain a threshold comparison result, and marking time points exceeding the preset alarm threshold as fault warning states based on the threshold comparison result to obtain a fault prediction result, includes: The abnormal score is compared with the preset alarm threshold to obtain a single-point comparison result. The single-point comparison result is then subjected to continuous detection according to a time window to obtain a continuous threshold count. The continuous threshold count is then compared with a preset continuous count threshold to obtain a threshold comparison result. Based on the threshold comparison results, the time points when the number of consecutive exceedances reaches a preset consecutive number threshold are marked as fault warning states, and the time points when the number of consecutive exceedances does not reach the preset consecutive number threshold are marked as normal operation states, thus obtaining a state marking sequence. The state marking sequence is then associated and bound with the corresponding timestamps to obtain fault prediction results.

7. The AI-based SMT production line equipment fault prediction method according to claim 1, characterized in that, The step of associating and encapsulating the fault prediction result with the corresponding equipment identifier and timestamp, and then outputting it to the production line operation and maintenance system includes: The set of time points marked as fault warning status is extracted from the fault prediction results to obtain the warning time point set. The warning time point set is associated and matched with the corresponding device identifier to obtain the device warning association record. The device warning association record is assembled with the corresponding timestamp and abnormal score value to obtain the warning information data packet. The early warning information data packet is encoded according to a preset data transmission protocol to obtain a standardized transmission message, and the standardized transmission message is sent to the production line operation and maintenance system through a data communication interface.

8. An AI-based SMT production line equipment fault prediction system, characterized in that, The system includes: The production line data acquisition module is used to collect equipment operating parameters from the chip mounter, reflow oven and printer of the surface mount technology production line, and add timestamps and equipment identifiers to obtain the original operating data set. The original operating data set is filled with missing values ​​and marked with outliers according to a preset sampling period to obtain the cleaned data set. The cleaned data set is grouped by equipment identifier and sorted by timestamp to obtain the equipment time sequence data sequence. The prediction model construction module is used to perform statistical calculations on the device time-series data sequence using a sliding window to obtain a time-domain feature matrix, send the time-domain feature matrix into the encoder network to perform feature encoding and reparameterized sampling to obtain a latent space vector, send the latent space vector into the decoder network to perform reconstruction to obtain reconstructed data, calculate the reconstruction error value based on the difference between the reconstructed data and the time-domain feature matrix, and compare the reconstruction error value with the error distribution parameters determined in the training phase to obtain an anomaly score. The equipment fault prediction module is used to perform a numerical comparison between the anomaly score and a preset alarm threshold to obtain a threshold comparison result, mark the time point that exceeds the preset alarm threshold as a fault warning state based on the threshold comparison result to obtain a fault prediction result, associate and encapsulate the fault prediction result with the corresponding equipment identifier and timestamp, and output it to the production line operation and maintenance system.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the AI-based SMT production line equipment fault prediction method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the AI-based SMT production line equipment fault prediction method according to any one of claims 1 to 7.

Images