A test data monitoring method for tablet mainboard production

Abstract

This invention belongs to the field of industrial data acquisition technology, specifically relating to a test data monitoring method for tablet motherboard production. The method includes: reading encoder pulse signals sent by the conveyor belt programmable logic controller during the reflow soldering process of the tablet motherboard, calculating real-time spatial coordinates, and generating a position timestamp; using the position timestamp as a reference anchor point, resampling and aligning the high-frequency temperature sequence and image data packets along the time axis with spatial coordinate mapping to generate a multi-source associated data frame with fused spatial location labels; inputting the multi-source associated data frame into a temporal feature extraction network containing a one-dimensional dilated causal convolutional layer and a max-pooling layer to extract local spatiotemporal features, dynamically adjusting the subsequent sampling frequency based on the absolute value of the feature gradient, and transmitting it to a monitoring server for anomaly detection and quality monitoring. This invention eliminates the time offset and spatial misalignment caused by asynchronous transmission of multi-source heterogeneous data.

Description

Technical Field

[0001] This invention belongs to the field of industrial data acquisition technology, specifically relating to a test data monitoring method for tablet computer motherboard production. Background Technology

[0002] In the reflow soldering process of flat panel motherboard manufacturing, it is necessary to collect and monitor data on the motherboard's temperature status and surface solder joint morphology. Current conventional industrial data acquisition solutions employ an architecture of independent operation and asynchronous transmission of multiple source devices. Specifically, a temperature sensor array sends high-frequency temperature sequences to a central server at a fixed sampling period, while an automated optical inspection device captures image data packets of the motherboard at a specific workstation and sends them independently. After receiving these two types of data, the central server performs simple time alignment based on the received timestamp of the data packet or the device's local timestamp. Subsequently, it inputs the temperature data and image data into a monitoring model to perform reflow soldering temperature curve anomaly detection and soldering quality monitoring.

[0003] In the aforementioned existing technical solutions, the temperature sensor array and the automated optical inspection equipment are located in different physical positions, and the data experiences varying network latency and queuing times during transmission to the central server. The simple alignment based on device timestamps fails to consider the actual physical movement trajectory of the motherboard on the conveyor belt, resulting in misalignment between the temperature data and image data received by the central server in terms of time and spatial coordinates. This spatiotemporal misalignment of multi-source heterogeneous data causes the monitoring model to incorrectly couple temperature features from different locations or at different times with image features. This leads to the monitoring system's judgments of abnormal reflow soldering temperature fluctuations and welding defects deviating from the actual physical state, resulting in distorted condition monitoring. Summary of the Invention

[0004] The purpose of this invention is to provide a test data monitoring method for tablet computer motherboard production, which can solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A test data monitoring method for tablet motherboard production includes: during the reflow soldering process of the tablet motherboard, a position resolution node reads the encoder pulse signal sent by the conveyor belt programmable logic controller in real time, calculates the real-time spatial coordinates of the tablet motherboard in combination with the physical topology of the production line, and generates a position timestamp; an edge alignment gateway synchronously receives a high-frequency temperature sequence from a temperature sensor array and an image data packet from an automatic optical inspection device, and performs time axis resampling and spatial coordinate mapping alignment on the high-frequency temperature sequence and the image data packet using the position timestamp as a reference anchor point to generate a multi-source associated data frame with fused spatial position labels; the multi-source associated data frame is input into a temporal feature extraction network deployed on the edge alignment gateway to extract local spatiotemporal features, the temporal feature extraction network including a one-dimensional dilated causal convolutional layer and a max pooling layer; the sampling frequency of the subsequent temperature sensor array is dynamically adjusted according to the absolute value of the feature gradient of the local spatiotemporal features, and the multi-source associated data frame and the local spatiotemporal features are transmitted to a monitoring server for reflow soldering temperature curve anomaly judgment and motherboard soldering quality monitoring.

[0006] Preferably, the step of calculating the real-time spatial coordinates of the flat panel motherboard and generating a position timestamp based on the production line physical topology includes: the position resolution node acquiring the segment physical length parameters of the conveyor belt, the pulley diameter parameters, and the total number of single-turn pulses of the encoder pulse signal; constructing a production line physical topology map containing multiple segment nodes based on the segment physical length parameters, and converting the encoder pulse signal into the moving distance of the flat panel motherboard on the conveyor belt; when the moving distance crosses a segment node in the production line physical topology map, the position resolution node generates the position timestamp containing the segment identifier by interpolation calculation based on the system clock time at the crossing time, the moving distance, and the absolute physical coordinates of the segment node.

[0007] Preferably, the step of using the location timestamp as a reference anchor point to perform time axis resampling and spatial coordinate mapping alignment on the high-frequency temperature sequence and the image data packet includes: the edge alignment gateway using the location timestamp as a node of a reference time grid, downsampling and interpolating the high-frequency temperature sequence according to the reference time grid to obtain an aligned temperature sequence; the edge alignment gateway extracting the shooting timestamp carried in the image data packet, calculating the time difference between the shooting timestamp and the location timestamp, calculating a spatial offset compensation amount based on the time difference and the running speed of the conveyor belt; using the spatial offset compensation amount to perform a row-direction translation transformation on the image pixel matrix corresponding to the image data packet, and concatenating the translated image pixel matrix with the aligned temperature sequence according to the location timestamp in the time dimension to generate the multi-source associated data frame.

[0008] Preferably, the step of inputting the multi-source associated data frame into the temporal feature extraction network deployed on the edge alignment gateway to extract local spatiotemporal features includes: the temporal feature extraction network performing channel dimension separation on the multi-source associated data frame, inputting the aligned temperature sequence into the one-dimensional dilated causal convolutional layer, and inputting the translated image pixel matrix into a two-dimensional spatial convolutional layer; the one-dimensional dilated causal convolutional layer contains multiple parallel dilated convolutional kernels, the dilation rate of the multiple parallel dilated convolutional kernels is set according to a geometric sequence, the one-dimensional dilated causal convolutional layer performs multi-scale receptive field convolution operation on the aligned temperature sequence, and outputs a multi-scale temperature feature map; the max pooling layer performs a fixed-window dimensionality reduction pooling operation on the multi-scale temperature feature map, and the output result after the dimensionality reduction pooling operation is cascaded and fused with the output result of the two-dimensional spatial convolutional layer to generate the local spatiotemporal features.

[0009] Preferably, the step of dynamically adjusting the sampling frequency of the subsequent temperature sensor array based on the absolute value of the feature gradient of the local spatiotemporal features includes: calculating the backpropagation gradient of the local spatiotemporal features in the time dimension, and extracting the temperature channel gradient vector from the backpropagation gradient; using the L2 norm of the temperature channel gradient vector as the absolute value of the feature gradient, and inputting the absolute value of the feature gradient into a finite state machine; the finite state machine is preset with a high-frequency sampling state, a mid-frequency sampling state, and a low-frequency sampling state; when the absolute value of the feature gradient falls into a first gradient interval, the current state is maintained; when the absolute value of the feature gradient falls into a second gradient interval higher than the first gradient interval, the state transitions to a high sampling frequency state and outputs a first frequency control word; when the absolute value falls into a third gradient interval lower than the first gradient interval, the state transitions to a low sampling frequency state and outputs a second frequency control word; the edge alignment gateway sends the first frequency control word or the second frequency control word to the temperature sensor array.

[0010] Preferably, the step of transmitting the multi-source associated data frame and the local spatiotemporal features to the monitoring server for reflow soldering temperature curve anomaly determination and motherboard soldering quality monitoring includes: the monitoring server reorganizing the multi-source associated data frame according to the fused spatial location label, and mapping the temperature data and image data in the reorganized multi-source associated data frame to node features and edge features in a graph structure, respectively; the monitoring server using a graph attention network to perform feature aggregation calculation on the node features and edge features in the graph structure to generate a graph feature containing global spatial topological relationships; the monitoring server performing cosine similarity calculation between the graph feature and the standard motherboard reflow soldering temperature curve feature, and when the cosine similarity is lower than a preset similarity threshold, generating a reflow soldering temperature curve anomaly determination result, and extracting image data within the time period corresponding to the anomaly determination result for soldering quality monitoring.

[0011] Preferably, before the position resolution node obtains the segmented physical length parameters of the conveyor belt, the method further includes: setting a photoelectric trigger switch at the physical location of each segment node of the conveyor belt, and attaching a uniquely identified RFID tag to the flat panel; when the flat panel passes the photoelectric trigger switch, the photoelectric trigger switch generates a trigger pulse signal, the position resolution node reads the unique identifier of the RFID tag and binds and stores the timestamp of the trigger pulse signal with the unique identifier; the position resolution node calculates the real-time running speed of the conveyor belt based on the time interval between the trigger pulse signals of two adjacent photoelectric trigger switches and the segmented physical length parameters, and uses the real-time running speed to perform online calibration of the pulse equivalent of the encoder pulse signal.

[0012] Preferably, the step of performing row-direction translation transformation on the image pixel matrix corresponding to the image data packet using the spatial offset compensation amount includes: converting the spatial offset compensation amount into the number of pixel rows to be translated in the image pixel matrix; when the number of pixel rows is not an integer, performing fractional pixel interpolation calculation on each row of pixels in the image pixel matrix to generate a sub-pixel-level translated image matrix; during the row-direction translation of the sub-pixel-level translated image matrix, for the empty pixel regions generated after translation, the edge alignment gateway extracts the edge pixel values ​​of the adjacent pixel rows of the empty pixel regions before translation, and uses a mirror symmetric copying strategy to fill the empty pixel regions with the edge pixel values ​​to generate the translated image pixel matrix without gaps.

[0013] Preferably, the step of concatenating and fusing the output of the dimensionality reduction pooling operation with the output of the two-dimensional spatial convolutional layer in a feature map concatenation includes: the edge alignment gateway copying and expanding the output of the dimensionality reduction pooling operation in the channel dimension, so that the number of channels in the expanded output is the same as the number of channels in the output of the two-dimensional spatial convolutional layer; concatenating the expanded output with the output of the two-dimensional spatial convolutional layer in the channel dimension to generate a concatenated feature map; inputting the concatenated feature map into a channel attention calculation layer, the channel attention calculation layer performing global average pooling and fully connected mapping on each channel of the concatenated feature map, and outputting channel weight coefficients; and the edge alignment gateway performing channel-by-channel multiplication operations on the channel weight coefficients and the concatenated feature map to generate the local spatiotemporal features.

[0014] Preferably, before the finite state machine transitions to a higher sampling frequency state and outputs a first frequency control word, or transitions to a lower sampling frequency state and outputs a second frequency control word, the method further includes: the finite state machine starting a de-jitter timer, continuously receiving the absolute value of the feature gradient within the timing period of the de-jitter timer; when the absolute value of the feature gradient received within the timing period of the de-jitter timer continuously falls within the second gradient interval or the third gradient interval, the finite state machine confirms the state transition is valid and outputs the first frequency control word or the second frequency control word; when the absolute value of the feature gradient received within the timing period of the de-jitter timer undergoes an interval jump, the finite state machine clears the de-jitter timer and maintains the current state unchanged.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention reads encoder pulse signals sent by the programmable logic controller (PLC) through a location resolution node and generates a location timestamp based on the production line's physical topology. At the edge alignment gateway, using this timestamp as a reference anchor point, time-axis resampling and spatial coordinate mapping alignment are performed on high-frequency temperature sequences and image data packets to generate multi-source associated data frames with fused spatial location tags. This process forcibly establishes a unified spatiotemporal reference between temperature data and image data at the edge, eliminating time offset errors and spatial misalignments caused by independent asynchronous transmission of multi-source heterogeneous data. This ensures a strict correspondence between the temperature features and image features input to the monitoring server in terms of physical location and occurrence time, and eliminates status monitoring distortion caused by feature miscoupling.

[0016] 2. This invention obtains local spatiotemporal features by inputting multi-source correlated data frames into a temporal feature extraction network containing one-dimensional dilated causal convolutional layers and max-pooling layers. Based on the absolute value of the feature gradient of these local spatiotemporal features, the sampling frequency of the subsequent temperature sensor array is dynamically adjusted. This allows the data acquisition frequency to adaptively adjust to the feature changes of the motherboard's actual operating conditions, reducing the amount of invalid redundant data transmission generated during the stable operation phase of the production line. Furthermore, by combining photoelectric trigger switches and RFID tags for online calibration of the encoder pulse signal pulse equivalent, and employing fractional-order pixel interpolation calculation and a mirror-symmetric replication strategy to perform row-direction translation of the image pixel matrix, the physical position calculation errors introduced by conveyor belt speed fluctuations and the interference from missing pixel regions caused by image translation are eliminated, ensuring the underlying data quality of multi-source correlated data frames with fused spatial location tags. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the overall process of industrial data acquisition and monitoring for tablet motherboard production according to the present invention. Figure 2This is a flowchart of the position analysis node coordinate calculation and pulse equivalent online calibration process of the present invention; Figure 3 This is a flowchart of the multi-source data time axis resampling and spatial coordinate mapping alignment process of the present invention; Figure 4 This is a flowchart of the local spatiotemporal feature extraction process of the temporal feature extraction network of the present invention; Figure 5 This is a flowchart of the sampling frequency dynamic adjustment based on feature gradient and finite state machine of the present invention; Figure 6 This is a flowchart for the monitoring server's reflow soldering temperature curve anomaly determination and welding quality monitoring of the present invention. Detailed Implementation

[0018] refer to Figure 1 In one embodiment, the reflow soldering process area of ​​the flat panel motherboard production line is equipped with a position resolution node, an edge alignment gateway, a temperature sensor array, an automated optical inspection device, and a monitoring server. The position resolution node establishes a wired communication connection with the programmable logic controller of the conveyor belt, and the edge alignment gateway establishes communication connections with the position resolution node, the temperature sensor array, the automated optical inspection device, and the monitoring server, respectively. The temperature sensor array is arranged at equal intervals along the running direction of the conveyor belt at the top of the internal cavity of the reflow oven, and the automated optical inspection device is located at the exit station of the reflow oven.

[0019] When the flat panel motherboard passes through the reflow soldering process, the position resolution node reads the encoder pulse signals sent by the conveyor belt's programmable logic controller in real time. The encoder is mounted on the drive pulley shaft of the conveyor belt and rotates synchronously with it, outputting a fixed number of pulse signals per revolution. The position resolution node counts the received encoder pulse signals, calculates the real-time spatial coordinates of the flat panel motherboard based on the production line physical topology, and generates a position timestamp. The production line physical topology is pre-stored in the position resolution node's local memory, containing the physical length, start coordinates, and end coordinates of each segment of the conveyor belt within the reflow soldering area. The position resolution node calculates the distance the flat panel motherboard moves on the conveyor belt based on the cumulative count of the encoder pulse signals, and, combined with the segment coordinate information in the production line physical topology, calculates the real-time spatial coordinates of the flat panel motherboard through linear interpolation. Simultaneously with generating the real-time spatial coordinates, the position resolution node reads the current system clock time, binds the real-time spatial coordinates to the current time, and generates a position timestamp containing both spatial coordinate and time information. The position resolution node then sends the generated position timestamp to the edge alignment gateway in real time.

[0020] The edge alignment gateway synchronously receives high-frequency temperature sequences from a temperature sensor array and image data packets from an automated optical inspection device. Each temperature sensor in the temperature sensor array collects temperature data inside the reflow oven at a preset initial sampling frequency, generating a high-frequency temperature sequence containing temperature values ​​and local sensor timestamps, and sends this sequence to the edge alignment gateway. The automated optical inspection device captures images of the flat panel motherboard's surface as it passes its inspection station, generating image data packets containing an image pixel matrix and a capture timestamp, and sends these packets to the edge alignment gateway. The edge alignment gateway uses the received location timestamp as a reference anchor point to resample the high-frequency temperature sequence and image data packets along the time axis and align them with spatial coordinates. The edge alignment gateway first uses the location timestamp as a unified time reference to resample the high-frequency temperature sequence along the time axis, ensuring that the time points of the resampled temperature data are completely consistent with the time points of the location timestamps. The edge alignment gateway then calculates the spatial offset of the tablet motherboard between the capture time and the corresponding location timestamp based on the time difference between the capture timestamp and the location timestamp in the image data packet. This spatial offset is used to perform a spatial coordinate mapping transformation on the image pixel matrix, ensuring that the physical location of the transformed image pixel matrix perfectly matches the physical location of the tablet motherboard corresponding to the location timestamp. The edge alignment gateway then associates the resampled temperature data with the spatially mapped image data according to the location timestamp, generating a multi-source associated data frame with fused spatial location labels. This multi-source associated data frame contains location timestamp information, spatial location label information, aligned temperature data, and aligned image data.

[0021] The edge alignment gateway inputs the generated multi-source correlated data frames into a locally deployed temporal feature extraction network to extract local spatiotemporal features from the data frames. The temporal feature extraction network comprises a one-dimensional dilated causal convolutional layer and a max-pooling layer. The one-dimensional dilated causal convolutional layer extracts the temporal features of the temperature data, while the max-pooling layer performs dimensionality reduction on the extracted features. The network preprocesses the multi-source correlated data frames, converting the temperature data into a one-dimensional temporal vector and the image data into a two-dimensional feature matrix. The one-dimensional temporal vector is input into the one-dimensional dilated causal convolutional layer for multi-scale temporal convolution operations, outputting a temperature temporal feature map. The temperature temporal feature map is then input into the max-pooling layer for dimensionality reduction pooling, outputting a dimensionality-reduced temperature feature vector. Finally, the network fuses the dimensionality-reduced temperature feature vector with the two-dimensional feature matrix of the image data to generate local spatiotemporal features containing both temperature temporal features and image spatial features.

[0022] The edge alignment gateway calculates the absolute value of the feature gradient of local spatiotemporal features and dynamically adjusts the sampling frequency of the subsequent temperature sensor array based on this value. The edge alignment gateway performs backpropagation calculations on the local spatiotemporal features to obtain the gradient value of the local spatiotemporal features relative to the input multi-source correlated data frame. The edge alignment gateway extracts the gradient component corresponding to the temperature data channel from the gradient value and calculates the absolute value of this gradient component as the absolute value of the feature gradient. The edge alignment gateway compares the absolute value of the feature gradient with several preset gradient thresholds and determines the target sampling frequency of the temperature sensor array based on the comparison results. When the absolute value of the feature gradient is greater than the first gradient threshold, the edge alignment gateway increases the sampling frequency of the temperature sensor array; when the absolute value of the feature gradient is less than the second gradient threshold, the edge alignment gateway decreases the sampling frequency of the temperature sensor array; when the absolute value of the feature gradient is between the first and second gradient thresholds, the edge alignment gateway maintains the current sampling frequency of the temperature sensor array. The edge alignment gateway sends a control command containing the target sampling frequency information to the temperature sensor array, and the temperature sensor array adjusts its own sampling frequency according to the received control command.

[0023] The edge alignment gateway transmits the generated multi-source correlated data frames and extracted local spatiotemporal features to the monitoring server via industrial Ethernet. The monitoring server receives the multi-source correlated data frames and local spatiotemporal features, and performs anomaly detection of the reflow soldering temperature curve and motherboard soldering quality monitoring. First, based on the spatial location tags in the multi-source correlated data frames, the monitoring server reassembles multi-source correlated data frames collected from different locations on the same motherboard, generating a complete reflow soldering process data sequence for the motherboard. The monitoring server extracts the temperature data from the complete reflow soldering process data sequence to generate the actual reflow soldering temperature curve of the motherboard. The monitoring server compares the actual reflow soldering temperature curve with a pre-stored standard reflow soldering temperature curve, calculating the similarity between the two curves. When the similarity is lower than a preset similarity threshold, the monitoring server determines that the reflow soldering temperature curve of the motherboard is abnormal. After determining that the reflow soldering temperature curve is abnormal, the monitoring server extracts the corresponding image data within the abnormal time period, performs soldering defect detection analysis on the image data, and identifies the type and location of soldering defects on the surface of the motherboard. The monitoring server generates monitoring results that include information on temperature curve anomalies and welding defects, and stores the monitoring results in a local database.

[0024] Table 1. Definition of Multi-Source Association Data Frame Structure Frame header identifier unsigned integers 2 Used to identify the start of a multi-source associated data frame, with a fixed value of 0xAA55. Frame sequence number unsigned integers 4 A unique sequence number for multi-source associated data frames, used for data frame order verification and loss detection. Location timestamp 64-bit floating-point number 8 System time measured in seconds, accurate to the microsecond level. Spatial location tags Structure 12 Includes the X-axis coordinates, Y-axis coordinates, and segment identifiers of the tablet motherboard, where the X-axis and Y-axis coordinates are 32-bit floating-point numbers, and the segment identifiers are 32-bit unsigned integers. Temperature data length unsigned integers 2 Byte length of subsequent temperature data fields Temperature data floating-point array variable The aligned temperature data consists of 32-bit floating-point numbers for each temperature value, and the array length is equal to the number of sensors in the temperature sensor array. Image data length unsigned integers 4 Byte length of subsequent image data fields Image data byte array variable The aligned image pixel matrix data is in grayscale image format, with each pixel value being an 8-bit unsigned integer. Frame checksum unsigned integers 2 The result of performing a CRC16 checksum calculation on all bytes of the entire data frame except for the frame header identifier and frame checksum. Frame end identifier unsigned integers 2 Used to identify the end of multi-source associated data frames, with a fixed value of 0x55AA. Table 1 defines the specific structure of a multi-source correlated data frame, containing 10 fields: frame header identifier, frame sequence number, location timestamp, spatial location tag, temperature data length, temperature data, image data length, image data, frame checksum, and frame trailer identifier. The data type, byte length, and specific meaning of each field are clearly defined. The frame header and trailer identifiers are used for data frame boundary identification; the frame sequence number is used for data frame order verification and loss detection; the location timestamp and spatial location tag are used to identify the physical location and time corresponding to the data frame; temperature data and image data are the core content of the multi-source correlated data frame; and the frame checksum is used for data frame transmission error detection.

[0025] In this embodiment, a unified location timestamp is generated by the location resolution node as a reference anchor point. Spatiotemporal alignment processing is performed on temperature and image data at the edge, eliminating time offset errors and spatial misalignment phenomena caused by independent asynchronous transmission of multi-source heterogeneous data. By deploying a temporal feature extraction network on the edge alignment gateway to extract local spatiotemporal features, and dynamically adjusting the sampling frequency of the temperature sensor array based on the absolute value of the feature gradient, adaptive adjustment of the data acquisition frequency is achieved. The monitoring server performs anomaly detection of temperature curves and welding quality monitoring on the reconstructed multi-source correlated data frames, ensuring the accuracy of the monitoring results.

[0026] refer to Figure 2 In a preferred embodiment, the position resolution node acquires the segmented physical length parameters of the conveyor belt, the pulley diameter parameters, and the total number of pulses per revolution of the encoder pulse signal. The segmented physical length parameters, which are pre-input by production line management personnel into the local memory of the position resolution node, include the physical lengths of the multiple segments into which the conveyor belt within the reflow soldering process area is divided. The pulley diameter parameter is the nominal diameter of the drive pulley of the conveyor belt, and the total number of pulses per revolution of the encoder pulse signal is the number of pulse signals output by the encoder per revolution. The position resolution node constructs a production line physical topology map containing multiple segment nodes based on the segmented physical length parameters. The production line physical topology map is represented in the form of a directed graph, where each node represents a segment node of the conveyor belt, the edges between nodes represent conveyor belt segments between adjacent segments, and the weight of the edge is the physical length of the corresponding conveyor belt segment. Each segment node is pre-assigned a unique segment identifier and absolute physical coordinates, which are defined with the production line's reference origin as the reference point.

[0027] The position resolution node counts the received encoder pulse signals in real time and converts them into the distance the flat panel moves on the conveyor belt. The formula for calculating the distance moved by converting the encoder pulse signal is: in, This indicates the distance the flat panel motherboard travels on the conveyor belt. This represents the cumulative count value of the encoder pulse signal. This indicates the diameter of the drive pulley of the conveyor belt. This indicates the total number of pulses per revolution of the encoder pulse signal. Pi is a constant.

[0028] The location resolution node calculates the cumulative movement distance of the tablet motherboard in real time and compares it with the cumulative length of each segment node in the production line physical topology diagram. When the cumulative movement distance crosses a segment node in the production line physical topology diagram, the location resolution node records the system clock time at the moment of crossing. Based on the system clock time at the moment of crossing, the cumulative movement distance, and the absolute physical coordinates of the segment node, the location resolution node generates a location timestamp containing the segment identifier through interpolation calculation. The specific interpolation calculation process is as follows: The location resolution node first determines the segment where the tablet motherboard is currently located and obtains the absolute physical coordinates and cumulative length of the starting and ending nodes of the segment. The location resolution node calculates the relative movement distance of the tablet motherboard within the current segment. Based on the ratio of the relative movement distance to the total length of the current segment, the absolute physical coordinates of the starting and ending nodes are linearly interpolated to obtain the real-time spatial coordinates of the tablet motherboard. The location resolution node binds the real-time spatial coordinates, the segment identifier of the current segment, and the system clock time at the moment of crossing to generate a location timestamp.

[0029] Photoelectric trigger switches are installed at the physical locations of each section node on the conveyor belt, and unique RFID tags are affixed to the tablet motherboards. The transmitter and receiver of the photoelectric trigger switches are installed on opposite sides of the conveyor belt. When a tablet motherboard passes the photoelectric trigger switch, it blocks the light emitted by the transmitter, preventing the receiver from receiving the light and generating a trigger pulse signal. The RFID tags are passive RFID tags, pre-written with unique identification information to distinguish different tablet motherboards. A location resolution node is connected to an RFID reader / writer, with the reader / writer's antenna installed near the photoelectric trigger switches. When a tablet motherboard passes the photoelectric trigger switch, the switch generates a trigger pulse signal and sends it to the location resolution node. Upon receiving the trigger pulse signal, the location resolution node controls the RFID reader / writer to read the unique identifier of the RFID tag on the tablet motherboard. The location resolution node binds the timestamp of the trigger pulse signal with the read unique identifier and stores it in its local memory.

[0030] The position resolution node calculates the real-time operating speed of the conveyor belt based on the time interval between the trigger pulse signals of two adjacent photoelectric trigger switches and the segment physical length parameters. The formula for calculating the real-time operating speed of the conveyor belt is: in, This indicates the real-time operating speed of the conveyor belt. This indicates the physical length of the conveyor belt section between two adjacent photoelectric trigger switches. This indicates the timestamp of the previous photoelectric trigger switch generating the trigger pulse signal. This indicates the timestamp of the trigger pulse signal generated by the next photoelectric trigger switch.

[0031] The position resolution node uses the calculated real-time running speed to perform online calibration of the pulse equivalent of the encoder pulse signal. The pulse equivalent refers to the conveyor belt travel distance corresponding to each encoder pulse signal. The calibration calculation formula for the pulse equivalent is: in, This indicates the calibrated pulse equivalent. This indicates the real-time operating speed of the conveyor belt. Indicates the encoder's sampling period. This indicates the total number of pulses per revolution in the encoder pulse signal.

[0032] The location resolution node recalculates the movement distance of the tablet motherboard using the calibrated pulse equivalent, and updates the real-time spatial coordinates and location timestamp of the tablet motherboard.

[0033] refer to Figure 3 The edge alignment gateway uses the received location timestamps as nodes of a reference time grid. It then downsamples and interpolates the high-frequency temperature sequence according to the reference time grid to obtain the aligned temperature sequence. The edge alignment gateway first extracts the time values ​​of all location timestamps and arranges these values ​​in ascending order to form the reference time grid. The time interval of the reference time grid is equal to the time difference between two adjacent location timestamps. The edge alignment gateway compares the timestamp of each temperature data point in the high-frequency temperature sequence with the node time of the reference time grid to determine the temperature data point corresponding to each reference time grid node. When the time of a reference time grid node is exactly the same as the timestamp of a temperature data point, the temperature value of that temperature data point is directly used as the value of the aligned temperature sequence at that node. When the time of a reference time grid node is between the timestamps of two adjacent temperature data points, a linear interpolation method is used to calculate the value of the aligned temperature sequence at that node. The formula for linear interpolation is: in, Represents the reference time grid node time Alignment temperature value at the location, This represents the temperature value of the previous temperature data point. This represents the timestamp of the previous temperature data point. This indicates the temperature value of the next temperature data point. This represents the timestamp of the next temperature data point.

[0034] The edge alignment gateway extracts the capture timestamp carried in the image data packet and calculates the time difference between the capture timestamp and the corresponding location timestamp. The corresponding location timestamp refers to the location timestamp closest to the capture time of the image data packet. The formula for calculating the time difference is: in, Indicates the time difference. This indicates the timestamp of the image data packet being captured. This indicates the corresponding location timestamp.

[0035] The edge alignment gateway calculates the spatial offset compensation based on the time difference and the conveyor belt's operating speed. The formula for calculating the spatial offset compensation is: in, Indicates the amount of spatial offset compensation. This indicates the real-time operating speed of the conveyor belt. This represents the time difference.

[0036] The edge alignment gateway uses spatial offset compensation to perform row-direction translation on the image pixel matrix corresponding to the image data packet. The row direction of the image pixel matrix is ​​consistent with the running direction of the conveyor belt. The edge alignment gateway first converts the spatial offset compensation into the number of pixel rows in the image pixel matrix that need to be translated. The conversion formula for the number of pixel rows is: in, Indicates the number of pixel rows that need to be shifted. Indicates the amount of spatial offset compensation. This represents the physical distance corresponding to each pixel in the image pixel matrix.

[0037] When the number of pixel rows is an integer, the edge alignment gateway directly translates all rows of the image pixel matrix along the row direction. The pixel position. When the number of pixel rows is not an integer, the edge alignment gateway performs fractional pixel interpolation calculations on each row of pixels in the image pixel matrix to generate a sub-pixel level translated image matrix. The fractional pixel interpolation calculation uses the bilinear interpolation method, and the specific calculation formula is as follows: in, This indicates the translated image in coordinates. Pixel value at that location, Represents the decimal part of the number of pixel rows, Indicates the original image in coordinates Pixel value at that location, Indicates the original image in coordinates The pixel value at that location.

[0038] During the row-wise translation of the subpixel-level image matrix, for the missing pixel regions created after translation, the edge alignment gateway extracts the edge pixel values ​​of the adjacent pixel rows of the missing pixel regions before translation and uses a mirror-symmetric copying strategy to fill the missing pixel regions with these edge pixel values. Specifically, when the image pixel matrix is ​​translated to the right, missing pixel regions will be generated on the left side. The edge alignment gateway extracts the edge pixel values ​​of the leftmost row of the original image... The pixel values ​​are copied symmetrically from the left side and used to fill the missing pixel areas on the left. When the image pixel matrix is ​​shifted to the left, missing pixel areas will be generated on the right side. The edge alignment gateway extracts the rightmost pixel values ​​from the original image. The pixel values ​​in each row are mirrored and then used to fill the empty pixel areas on the right. The edge alignment gateway generates a translated image pixel matrix without any gaps.

[0039] The edge alignment gateway concatenates the translated image pixel matrix with the aligned temperature sequence according to the location timestamp along the time dimension to generate a multi-source associated data frame.

[0040] Table 2 Production Line Physical Topology Section Parameters Table 2 lists the detailed parameters of each section in the production line's physical topology, including section identifier, start node coordinates, end node coordinates, section length, photoelectric trigger switch number, and RFID reader / writer number. Each section is 1.200 meters long, and the Y-axis values ​​for both the start and end node coordinates are 0.500 meters, indicating that the conveyor belt runs in a straight line along the X-axis. Each section has a photoelectric trigger switch and an RFID reader / writer at its start node for generating trigger pulse signals and reading RFID tags.

[0041] In this embodiment, by constructing a physical topology map of the production line containing multiple segment nodes and combining it with encoder pulse signals to calculate the real-time spatial coordinates of the flat panel motherboard, the accuracy of position calculation is improved. Through the combined use of photoelectric trigger switches and RFID tags, online calibration of the encoder pulse signal pulse equivalent is achieved, eliminating physical position calculation errors introduced by conveyor belt speed fluctuations. By downsampling and interpolating the high-frequency temperature sequence and performing row-direction translation on the image pixel matrix, precise alignment of temperature data and image data on the time axis and spatial coordinates is achieved. By employing fractional-order pixel interpolation calculation and a mirror-symmetric replication strategy, the accuracy of image translation transformation is improved, and interference from missing pixel areas caused by image translation is eliminated.

[0042] refer to Figure 4In a preferred embodiment, the temporal feature extraction network performs channel dimension separation on multi-source correlated data frames, inputs the aligned temperature sequence into a one-dimensional dilated causal convolutional layer, and inputs the translated image pixel matrix into a two-dimensional spatial convolutional layer. The temporal feature extraction network first parses the multi-source correlated data frames, extracting the aligned temperature sequence and the translated image pixel matrix. The aligned temperature sequence is a one-dimensional vector, its length equal to the number of sensors in the temperature sensor array. The translated image pixel matrix is ​​a two-dimensional matrix, its number of rows and columns equal to the height and width of the image, respectively. The temporal feature extraction network transforms the aligned temperature sequence into a shape... The tensor, in which This indicates the length of the aligned temperature sequence. The temporal feature extraction network transforms the translated image pixel matrix into a shape of... The tensor, in which Indicates the height of the image. Indicates the width of the image.

[0043] The one-dimensional dilated causal convolutional layer contains multiple parallel dilated convolutional kernels, with the dilation rates of these kernels set according to a geometric progression. In this embodiment, the one-dimensional dilated causal convolutional layer contains four parallel dilated convolutional kernels with dilation rates of 1, 2, 4, and 8, forming a geometric progression with a common ratio of 2. Each dilated convolutional kernel has a size of 3, a stride of 1, and is padded causally. Causal padding ensures that the output of the convolution operation depends only on the input data from the current and previous time steps, without introducing information from future time steps. The one-dimensional dilated causal convolutional layer performs multi-scale receptive field convolution operations on the aligned temperature sequence, outputting a multi-scale temperature feature map. The formula for calculating the multi-scale temperature feature map is: in, Represents a multi-scale temperature feature map. This indicates that a splicing operation is performed along the channel dimension. Indicates the input temperature sequence Expansion rate One-dimensional dilated causal convolution operation.

[0044] The max pooling layer performs a fixed-window dimensionality reduction pooling operation on the multi-scale temperature feature map. The pooling window size of the max pooling layer is 2, the stride is 2, and the filling method is effective filling. The max pooling operation reduces the dimensionality of the multi-scale temperature feature map in the time dimension, retaining the maximum value within each pooling window. The calculation formula for the max pooling operation is: in, This represents the output result after the dimensionality reduction pooling operation. This indicates a max pooling operation.

[0045] The two-dimensional spatial convolutional layer extracts spatial features from the pixel matrix of the translated image. It consists of two consecutive two-dimensional convolutional layers, each with a kernel size of 3, a stride of 1, and the same padding method. The activation function is ReLU. The first two-dimensional convolutional layer has 16 output channels, and the second has 32. The output of the two-dimensional spatial convolutional layer is an image spatial feature map with the following shape: .

[0046] The edge alignment gateway replicates and expands the output of the dimensionality reduction pooling operation along the channel dimension, ensuring that the number of channels in the expanded output is the same as the number of channels in the output of the two-dimensional spatial convolutional layer. The calculation formula for the replication and expansion operation is as follows: in, This indicates the expanded output result. Indicates will Repeat 32 times along the channel dimension.

[0047] The edge alignment gateway concatenates the expanded output with the output of the 2D spatial convolutional layer along the channel dimension to generate a concatenated feature map. The formula for calculating the concatenated feature map is: in, This represents the splicing of feature maps. This represents the output of a two-dimensional spatial convolutional layer.

[0048] The edge alignment gateway inputs the stitched feature map into the channel attention computation layer. The channel attention computation layer performs global average pooling and fully connected mapping on each channel of the stitched feature map, outputting channel weight coefficients. The specific calculation process of the channel attention computation layer is as follows: First, a global average pooling operation is performed on the stitched feature map, compressing the two-dimensional feature map of each channel into a scalar value. The formula for calculating the global average pooling operation is: in, Indicates the first Global average pooling results for each channel. and These represent the height and width of the spliced ​​feature map, respectively. Indicates the spliced ​​feature map in coordinates First The characteristic values ​​of each channel.

[0049] The global average pooling result is then input into two consecutive fully connected layers. The first fully connected layer has 1 / 16 the number of channels and uses the ReLU activation function; the second fully connected layer has the same number of channels and uses the Sigmoid activation function. The formula for calculating the fully connected mapping is: in, Indicates the first The weighting coefficients of each channel This represents the Sigmoid activation function. and These represent the weight matrix and bias vector of the first fully connected layer, respectively. and These represent the weight matrix and bias vector of the second fully connected layer, respectively.

[0050] The edge alignment gateway performs channel-by-channel multiplication on the stitched feature map to generate local spatiotemporal features. The formula for channel-by-channel multiplication is as follows: in, Representing local spatiotemporal features in coordinates First The characteristic values ​​of each channel.

[0051] refer to Figure 5 The edge alignment gateway calculates the backpropagation gradient of local spatiotemporal features along the time dimension and extracts the temperature channel gradient vector from the backpropagation gradient. The edge alignment gateway inputs the local spatiotemporal features into a simple fully connected layer, whose output is a scalar value. The edge alignment gateway calculates the gradient of this scalar value relative to the input multi-source correlated data frame, obtaining the backpropagation gradient. The formula for calculating the backpropagation gradient is: in, This represents the backpropagation gradient. This represents the output scalar value of the fully connected layer. This indicates the input of a multi-source associated data frame.

[0052] The edge alignment gateway extracts the gradient components of the corresponding temperature data channels from the backpropagation gradient, forming a temperature channel gradient vector. The edge alignment gateway then calculates the L2 norm of the temperature channel gradient vector as the absolute value of the feature gradient. The formula for calculating the L2 norm is: in, This represents the L2 norm of the temperature channel gradient vector, i.e., the absolute value of the characteristic gradient. The first element representing the gradient vector of the temperature channel One element, This represents the length of the temperature channel gradient vector.

[0053] The edge alignment gateway inputs the absolute value of the feature gradient into a finite state machine. The finite state machine has pre-defined high-frequency, mid-frequency, and low-frequency sampling states. The sampling frequency for the high-frequency state is 100Hz, for the mid-frequency state it is 50Hz, and for the low-frequency state it is 10Hz. The finite state machine has three pre-defined gradient intervals: the first gradient interval is... The second gradient interval is The third gradient interval is .in, For low gradient threshold, The threshold is set to the high gradient. When the absolute value of the feature gradient falls within the first gradient interval, the finite state machine maintains its current state. When the absolute value of the feature gradient falls within the second gradient interval, the finite state machine transitions to a higher sampling frequency state and outputs the first frequency control word. When the absolute value of the feature gradient falls within the third gradient interval, the finite state machine transitions to a lower sampling frequency state and outputs the second frequency control word.

[0054] Before the finite state machine (FSM) transitions to a higher sampling frequency state and outputs the first frequency control word, or transitions to a lower sampling frequency state and outputs the second frequency control word, it starts a debounce timer. The debounce timer's timing period is preset to 100 milliseconds. During the debounce timer's timing period, the FSM continuously receives the absolute values ​​of the feature gradients. When all the absolute values ​​of the feature gradients received during the debounce timer's timing period continuously fall within the second or third gradient interval, the FSM confirms the state transition is valid and outputs the first or second frequency control word. When the absolute values ​​of the feature gradients received during the debounce timer's timing period undergo an interval jump, the FSM resets the debounce timer and maintains its current state.

[0055] The edge alignment gateway sends either a first frequency control word or a second frequency control word to the temperature sensor array. Each temperature sensor in the array adjusts its sampling frequency according to the received frequency control word.

[0056] refer to Figure 6The monitoring server reassembles multi-source correlated data frames based on their fused spatial location tags, mapping the temperature and image data within the reassembled frames to node and edge features in a graph structure, respectively. First, the monitoring server groups all multi-source correlated data frames belonging to the same motherboard based on their unique RFID tags. Then, based on the spatial location tags in each group, the server arranges the multi-source correlated data frames in chronological order, generating a complete reflow soldering process data sequence for the motherboard. The monitoring server maps each multi-source correlated data frame in the complete reflow soldering process data sequence to a node in the graph structure, with the node's feature vector being the aligned temperature sequence within that multi-source correlated data frame. Finally, the monitoring server maps the connection between two adjacent nodes to an edge in the graph structure, with the edge's feature vector being the feature vector of the image data from the corresponding multi-source correlated data frames.

[0057] The monitoring server utilizes a graph attention network to perform feature aggregation calculations on node and edge features in the graph structure, generating graph features that include global spatial topological relationships. The graph attention network consists of two graph attention layers, each with 8 attention heads and 64 output channels. The formula for calculating the attention coefficient of the graph attention layer is as follows: in, Represents a node For nodes Attention coefficient This represents the LeakyReLU activation function. The weight vector representing the attention mechanism. The linear transformation matrix representing the node features. and Representing nodes respectively and nodes eigenvectors, Represents a node and nodes The edge feature vectors between them This represents a vector concatenation operation. Represents a node The set of neighboring nodes.

[0058] The formula for calculating the output of the graph attention layer is: in, Nodes representing the output of the graph attention layer eigenvectors, This represents the ELU activation function.

[0059] The monitoring server concatenates all node feature vectors output by the graph attention network to generate a graph feature containing global spatial topological relationships.

[0060] The monitoring server calculates the cosine similarity between the spectral features and the reflow temperature profile features of a standard motherboard. The formula for calculating the cosine similarity is: in, Represents cosine similarity. This represents the spectral features of the motherboard of the tablet to be tested. This indicates the characteristics of the standard motherboard reflow soldering temperature profile.

[0061] When the cosine similarity is lower than a preset similarity threshold, the monitoring server generates an anomaly determination result for the reflow soldering temperature curve. The monitoring server extracts image data within the time period corresponding to the anomaly determination result and inputs the image data into a pre-trained welding defect detection model for welding quality monitoring. The welding defect detection model uses the YOLOv5 target detection model, which can identify various common welding defect types such as cold welds, bridging welds, and missing welds.

[0062] Table 3. Finite State Machine State Transition Rules Table 3 details the state transition rules of the finite state machine, including five aspects: the current state, the absolute value range of the feature gradient, the debouncing timer state, the target state, and the output frequency control word. Before a state transition, the debouncing timer must be started. A state transition and the corresponding frequency control word will only be executed and the absolute value of the feature gradient output will only occur if the absolute value of the feature gradient continuously falls within the target range during the debouncing timer's timing period. If the absolute value of the feature gradient jumps within the range during the debouncing timer's timing period, the debouncing timer is cleared and the current state remains unchanged.

[0063] In this embodiment, multi-scale temporal features of temperature data are extracted using a one-dimensional dilated causal convolutional layer containing multiple parallel dilated convolutional kernels, improving the ability to extract temporal features. Channel attention computation layers are used to adjust channel weights in the stitched feature map, enhancing the expressive power of important features. A finite state machine combined with a de-jitter timer is used to dynamically adjust the sampling frequency of the temperature sensor array, avoiding frequent fluctuations in the sampling frequency. A graph attention network is used to perform feature aggregation computation on the recombined multi-source associated data frames, generating a graph feature containing global spatial topological relationships, improving the accuracy of anomaly detection in reflow soldering temperature curves.

Claims

1. A test data monitoring method for tablet mainboard production, characterized in that, The method includes: when the flat motherboard undergoes the reflow soldering process, the position resolution node reads the encoder pulse signal sent by the conveyor belt programmable logic controller in real time, calculates the real-time spatial coordinates of the flat motherboard in combination with the physical topology of the production line, and generates a position timestamp. The edge alignment gateway synchronously receives high-frequency temperature sequences from a temperature sensor array and image data packets from an automated optical inspection device. Using the location timestamp as a reference anchor point, it performs time axis resampling and spatial coordinate mapping alignment on the high-frequency temperature sequences and the image data packets to generate multi-source associated data frames with fused spatial location labels. The multi-source associated data frame is input into the temporal feature extraction network deployed on the edge alignment gateway to extract local spatiotemporal features. The temporal feature extraction network includes a one-dimensional dilated causal convolutional layer and a max pooling layer. The sampling frequency of the subsequent temperature sensor array is dynamically adjusted based on the absolute value of the feature gradient of the local spatiotemporal features, and the multi-source associated data frame and the local spatiotemporal features are transmitted to the monitoring server for reflow soldering temperature curve anomaly determination and motherboard soldering quality monitoring.

2. The method for monitoring test data for production of a tablet motherboard according to claim 1, wherein, The step of calculating the real-time spatial coordinates of the flat motherboard and generating a position timestamp by combining the physical topology of the production line includes: the position parsing node obtaining the segmented physical length parameters of the conveyor belt, the pulley diameter parameters, and the total number of single-turn pulses of the encoder pulse signal; Based on the segmented physical length parameters, a production line physical topology map containing multiple segment nodes is constructed, and the encoder pulse signal is converted into the moving distance of the flat motherboard on the conveyor belt; When the movement distance crosses a segment node in the physical topology diagram of the production line, the location resolution node generates a location timestamp containing the segment identifier by interpolation calculation based on the system clock time at the crossing time, the movement distance, and the absolute physical coordinates of the segment node.

3. The method for monitoring test data for tablet motherboard production of claim 1, wherein, The step of using the location timestamp as a reference anchor point to perform time axis resampling and spatial coordinate mapping alignment on the high-frequency temperature sequence and the image data packet includes: the edge alignment gateway uses the location timestamp as a node of the reference time grid, and performs downsampling interpolation on the high-frequency temperature sequence according to the reference time grid to obtain the aligned temperature sequence; The edge alignment gateway extracts the shooting timestamp carried in the image data packet, calculates the time difference between the shooting timestamp and the location timestamp, and calculates the spatial offset compensation amount based on the time difference and the running speed of the conveyor belt. The image pixel matrix corresponding to the image data packet is transformed by row direction using the spatial offset compensation amount, and the transformed image pixel matrix is ​​then concatenated with the aligned temperature sequence according to the time dimension of the location timestamp to generate the multi-source associated data frame.

4. The method for monitoring test data for tablet motherboard production of claim 1, wherein, The step of inputting the multi-source associated data frame into the temporal feature extraction network deployed on the edge alignment gateway to extract local spatiotemporal features includes: the temporal feature extraction network performing channel dimension separation on the multi-source associated data frame, inputting the aligned temperature sequence into the one-dimensional dilated causal convolutional layer, and inputting the translated image pixel matrix into the two-dimensional spatial convolutional layer. The one-dimensional dilated causal convolutional layer contains multiple parallel dilated convolutional kernels, and the dilation rate of the multiple parallel dilated convolutional kernels is set according to a geometric sequence. The one-dimensional dilated causal convolutional layer performs multi-scale receptive field convolution operation on the aligned temperature sequence and outputs a multi-scale temperature feature map. The max pooling layer performs a fixed-window dimensionality reduction pooling operation on the multi-scale temperature feature map, and then concatenates and fuses the output of the dimensionality reduction pooling operation with the output of the two-dimensional spatial convolutional layer to generate the local spatiotemporal features.

5. The test data monitoring method for tablet computer motherboard production according to claim 1, characterized in that, The step of dynamically adjusting the sampling frequency of the subsequent temperature sensor array based on the absolute value of the feature gradient of the local spatiotemporal features includes: calculating the backpropagation gradient of the local spatiotemporal features in the time dimension, and extracting the temperature channel gradient vector in the backpropagation gradient. The L2 norm of the temperature channel gradient vector is used as the absolute value of the feature gradient, and the absolute value of the feature gradient is input into the finite state machine. The finite state machine is preset with a high-frequency sampling state, a medium-frequency sampling state, and a low-frequency sampling state. When the absolute value of the feature gradient falls into the first gradient interval, the current state is maintained. When the absolute value of the feature gradient falls into the second gradient interval which is higher than the first gradient interval, the machine transitions to the high sampling frequency state and outputs the first frequency control word. When the absolute value of the feature gradient falls into the third gradient interval which is lower than the first gradient interval, the machine transitions to the low sampling frequency state and outputs the second frequency control word. The edge-aligned gateway sends the first frequency control word or the second frequency control word to the temperature sensor array.

6. The test data monitoring method for tablet computer motherboard production according to claim 1, characterized in that, The step of transmitting the multi-source associated data frame and the local spatiotemporal features to the monitoring server for reflow soldering temperature curve anomaly determination and motherboard soldering quality monitoring includes: the monitoring server reorganizes the multi-source associated data frame according to the fused spatial location label, and maps the temperature data and image data in the reorganized multi-source associated data frame to node features and edge features in a graph structure, respectively. The monitoring server uses a graph attention network to perform feature aggregation calculations on the node features and edge features in the graph structure, generating graph features that contain global spatial topological relationships; The monitoring server calculates the cosine similarity between the spectral features and the standard motherboard reflow soldering temperature curve features. When the cosine similarity is lower than a preset similarity threshold, an anomaly judgment result for the reflow soldering temperature curve is generated, and image data within the time period corresponding to the anomaly judgment result is extracted for welding quality monitoring.

7. The test data monitoring method for tablet computer motherboard production according to claim 2, characterized in that, Before the location resolution node obtains the segmented physical length parameters of the conveyor belt, the method further includes: setting a photoelectric trigger switch at the physical location of each segment node of the conveyor belt, and attaching a uniquely identified radio frequency tag to the flat panel motherboard; When the tablet motherboard passes the photoelectric trigger switch, the photoelectric trigger switch generates a trigger pulse signal. The location resolution node reads the unique identifier of the RFID tag and binds and stores the timestamp of the trigger pulse signal with the unique identifier. The position resolution node calculates the real-time running speed of the conveyor belt based on the time interval between the trigger pulse signals of two adjacent photoelectric trigger switches and the segment physical length parameter, and uses the real-time running speed to perform online calibration of the pulse equivalent of the encoder pulse signal.

8. The test data monitoring method for tablet computer motherboard production according to claim 3, characterized in that, The step of performing row-direction translation transformation on the image pixel matrix corresponding to the image data packet using the spatial offset compensation amount includes: converting the spatial offset compensation amount into the number of pixel rows in the image pixel matrix that need to be translated; When the number of pixel rows is not an integer, fractional pixel interpolation is performed on each row of pixels in the image pixel matrix to generate a sub-pixel level translation image matrix. During the row-direction translation of the subpixel-level translated image matrix, for the missing pixel regions generated after translation, the edge alignment gateway extracts the edge pixel values ​​of the adjacent pixel rows of the missing pixel regions before translation, and uses a mirror symmetric copying strategy to fill the missing pixel regions with the edge pixel values, generating the translated image pixel matrix without gaps.

9. The test data monitoring method for tablet computer motherboard production according to claim 4, characterized in that, The step of performing feature map concatenation and fusion of the output result after dimensionality reduction pooling operation and the output result of the two-dimensional spatial convolutional layer includes: the edge alignment gateway copies and expands the output result after dimensionality reduction pooling operation in the channel dimension, so that the number of channels of the expanded output result is the same as the number of channels of the output result of the two-dimensional spatial convolutional layer. The expanded output is concatenated with the output of the two-dimensional spatial convolutional layer along the channel dimension to generate a concatenated feature map. The stitched feature map is input into the channel attention calculation layer, which performs global average pooling and fully connected mapping on each channel of the stitched feature map and outputs channel weight coefficients. The edge alignment gateway performs a channel-by-channel multiplication operation between the channel weight coefficients and the stitched feature map to generate the local spatiotemporal features.

10. The test data monitoring method for tablet computer motherboard production according to claim 5, characterized in that, Before the finite state machine transitions to a high sampling frequency state and outputs a first frequency control word, or transitions to a low sampling frequency state and outputs a second frequency control word, the method further includes: the finite state machine starts a debounce timer and continuously receives the absolute value of the feature gradient during the timing period of the debounce timer; When the absolute value of the characteristic gradient received within the timing period of the anti-shake timer continuously falls into the second gradient interval or the third gradient interval, the finite state machine confirms that the state transition is valid and outputs the first frequency control word or the second frequency control word. When the absolute value of the feature gradient received within the timing period of the anti-shake timer changes within an interval, the finite state machine clears the anti-shake timer and maintains the current state unchanged.

Images