A defect tracing method and system for PCBA surface mount lines

Abstract

The application relates to the field of PCBA surface mounting technology, in particular to a defect tracing method and system for a PCBA surface mounting line, which comprises the following steps: when a cloud platform receives defect types of a target PCB board fed back by a PCBA surface mounting line, cross-station characteristic data of the target PCB board is acquired based on a unique identifier of the target PCB board; a historical sample set is acquired based on the defect types of the target PCB board and a pre-constructed historical cross-station characteristic database; the causal contribution degree of each station in the PCBA surface mounting line to the defects of the target PCB board is acquired based on the cross-station characteristic data of the target PCB board and the historical sample set and a causal contribution degree calculation algorithm; and the defects of the target PCB board are traced to the stations based on the causal contribution degree of each station in the PCBA surface mounting line to the defects of the target PCB board. The method significantly improves the accuracy and objectivity of defect tracing of the PCBA surface mounting line, and provides reliable data support for process rectification, parameter optimization and defect source treatment of the production line.

Description

Technical Field

[0001] This application relates to the field of PCBA surface mount technology, and in particular to a method and system for tracing defects in PCBA surface mount lines. Background Technology

[0002] PCBA surface mount production lines consist of multiple process stations sequentially, including solder paste printing, inspection, placement, reflow soldering, and post-reflow inspection. The process parameters at each station are easily affected by environmental fluctuations, equipment aging, and material differences, which can readily lead to various PCB board defects such as solder paste misalignment, placement misalignment, and cold solder joints. Current PCBA defect handling methods mostly involve simple post-defect reviews after detection, lacking cross-station correlation analysis capabilities and making it difficult to accurately pinpoint the true source station of the defect.

[0003] Currently, most conventional defect tracing methods rely solely on the detection data of a single workstation for independent identification, ignoring the process coupling and chain transmission effects between upstream and downstream workstations. This can easily lead to misjudging statistical correlations caused by process linkages as the root cause of defects, resulting in low tracing accuracy. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] In view of the above-mentioned shortcomings and deficiencies of the prior art, this application provides a method and system for tracing defects in PCBA surface mount lines, which solves the technical problem that the existing PCBA surface mount line defect tracing methods rely only on the detection data of a single workstation and ignore the process coupling relationship between upstream and downstream workstations, thus resulting in low tracing accuracy.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the main technical solutions adopted in this application include:

[0008] In a first aspect, embodiments of this application provide a defect tracing method for PCBA surface mount lines. The method is implemented based on a pre-deployed PCBA surface mount line and a cloud platform. The PCBA surface mount line includes multiple workstations arranged sequentially according to the process flow. The method includes: when the cloud platform receives a defect type of a target PCB board from the PCBA surface mount line, the cloud platform obtains cross-workstation feature data of the target PCB board during its flow through the PCBA surface mount line based on a pre-set unique identifier of the target PCB board; the cloud platform obtains a historical sample set based on the defect type of the target PCB board and a pre-built historical cross-workstation feature database; wherein the historical sample set includes historical data corresponding to multiple normal PCB boards. The cloud platform collects historical cross-station feature data and historical cross-station feature data corresponding to multiple defective PCBs with the same defect type as the target PCB. Based on the cross-station feature data of the target PCB and the historical sample set, as well as a pre-set causal contribution statistical algorithm, the cloud platform obtains the causal contribution of each station in the PCBA surface mounting line to the defect of the target PCB. The causal contribution statistical algorithm is used to determine the causal contribution of each station in the PCBA surface mounting line to the defect of the target PCB based on the distribution of the cross-station feature data of the target PCB in the historical sample set. Based on the causal contribution of each station in the PCBA surface mounting line to the defect of the target PCB, the cloud platform performs station-based source tracing of the defect of the target PCB.

[0009] Optionally, the PCBA surface mount line includes: a solder paste printing station, a solder paste printing inspection station, a chip mounting station, a chip mounting position inspection station, a reflow oven, and a soldering quality inspection station arranged sequentially according to the process flow; when the cloud platform receives the defect type of the target PCB board from the PCBA surface mount line, the cloud platform obtains cross-station characteristic data of the target PCB board during the PCBA surface mount line flow based on the unique identifier pre-set for the target PCB board, including: when the cloud platform receives the defect type of the target PCB board from the soldering quality inspection station, the cloud platform obtains the corresponding solder paste printing process data, solder paste printing inspection data, chip mounting station process data, chip mounting position inspection data, reflow oven temperature control process data, and soldering quality inspection data based on the unique identifier pre-set for the target PCB board; solder paste printing process data The data includes: process data from the solder paste printing station during solder paste printing on the target PCB board; solder paste printing inspection data from the solder paste printing inspection station; process data from the placement station during placement on the target PCB board; placement position detection data from the placement position detection station; reflow oven temperature control process data from the reflow oven during soldering on the target PCB board; and soldering quality inspection data from the soldering quality inspection station. The cloud platform extracts features from the solder paste printing process data, solder paste printing inspection data, placement station process data, placement position detection data, reflow oven temperature control process data, and soldering quality inspection data, and sorts the quantified features extracted from all data according to the process flow to obtain cross-station feature data of the target PCB board.

[0010] Optionally, the cross-station feature data of the target PCB board and each of the historical cross-station feature data includes multiple station feature types; the cloud platform obtains a historical sample set based on the defect type of the target PCB board and a pre-built historical cross-station feature database, including: the cloud platform, based on the defect type of the target PCB board, filters out historical cross-station feature data corresponding to all defective PCB boards with the same defect type as the target PCB board from the pre-built historical cross-station feature database, as well as historical cross-station feature data corresponding to all normal PCB boards without defects; and labels all defective PCB boards with defect tags; the cloud platform, based on a pre-set... The equal-frequency binning discretization algorithm is used to discretize each selected historical cross-station feature data. The equal-frequency binning discretization strategy includes: for each station feature type, statistically analyzing the feature values ​​corresponding to that station feature type in all historical cross-station feature data to obtain the feature value distribution corresponding to each station feature type; and based on the feature value distribution corresponding to each station feature type, dividing each station feature type into a preset number of discrete intervals so that the feature values ​​corresponding to each station feature type in all historical cross-station feature data are mapped to the discrete intervals; and constructing a historical sample set based on all historical cross-station feature data after discretization and all defect labels.

[0011] Optionally, the cloud platform, based on the cross-station feature data of the target PCB board and the historical sample set, and a pre-set causal contribution statistical algorithm, obtains the causal contribution of each station in the PCBA surface mounting line to the defects of the target PCB board. This includes: the cloud platform mapping the feature values ​​of each station feature type in the cross-station feature data of the target PCB board to the corresponding discrete interval; the cloud platform, based on all historical cross-station feature data and all defect type labels in the historical sample set, obtaining the joint probability distribution, station feature marginal distribution, and defect label marginal distribution corresponding to each station feature type in the historical sample set; wherein, the joint probability distribution is the historical cross-station feature data with / without defect labels in all historical cross-station feature data corresponding to the station feature type. The probability distribution of the feature values ​​corresponding to the data within each discrete interval; the marginal distribution of workstation features is the probability distribution of the feature values ​​of all historical cross-workstation feature data corresponding to the workstation feature type within each discrete interval; the marginal distribution of defect labels is the probability distribution of having defect labels or not having defects among all historical cross-workstation feature data corresponding to the workstation feature type; the cloud platform obtains the mutual information corresponding to each workstation feature type based on the joint probability distribution, normal joint probability distribution, workstation feature marginal distribution, and defect label marginal distribution corresponding to each workstation feature type in the historical sample set; the cloud platform obtains the causal contribution of each workstation in the PCBA surface mounting line to the defects of the target PCB board based on the cross-workstation feature data of the target PCB board and the mutual information corresponding to each workstation feature type.

[0012] Optionally, the cloud platform obtains mutual information corresponding to each workstation feature type based on the joint probability distribution, normal joint probability distribution, workstation feature marginal distribution, and defect label marginal distribution corresponding to each workstation feature type in the historical sample set. This includes: the cloud platform obtains mutual information between each workstation feature type and the defect type based on the joint probability distribution, workstation feature marginal distribution, and defect label marginal distribution corresponding to each workstation feature type in the historical sample set, and a pre-set formula (Formula 1); the formula (Formula 1) is:

[0013] ;

[0014] Where r(X) i Let p(k,d) be the mutual information between the workstation feature type and the defect type at index i, k be the index of the discrete interval, K be the number of discrete intervals, and d be the index of whether or not a defect label is present. p(k) is the probability that the feature value in the marginal distribution of the workstation feature type falls within the kth discrete interval, p(d) is the probability that the feature value in the marginal distribution of the defect label of the workstation feature type has a defect label / does not have a defect label, and p(k,d) is the probability that the feature value in the joint probability distribution of the workstation feature type has / does not have a defect label and is within the kth discrete interval.

[0015] Optionally, the cloud platform obtains the causal contribution of each workstation in the PCBA surface mounting line to the defects of the target PCB board based on the cross-workstation feature data of the target PCB board and the mutual information corresponding to each workstation feature type. This includes: the cloud platform normalizes the mutual information corresponding to all workstation feature types to obtain the basic causal contribution corresponding to each workstation feature type; wherein the sum of the basic causal contributions corresponding to all workstation feature types is 1; the cloud platform obtains the deviation weight corresponding to each workstation feature type based on the cross-workstation feature data corresponding to the target PCB board, the historical cross-workstation feature data corresponding to all normal PCB boards in the historical sample set, and a pre-set formula two; the formula two is:

[0016] ;

[0017] Among them, w i x is the deviation weight corresponding to the feature type of the workstation with index i. i,目标板 Let i be the feature value of the cross-station feature data corresponding to the target PCB board at the station feature type with index i. This represents the mean of the feature values ​​for all normal PCB boards at the workstation feature type with index i. The standard deviation of the feature values ​​of all normal boards on the feature type of workstation with index i; the cloud platform, based on the deviation weight corresponding to each feature type of workstation and the pre-set workstation feature table, performs a weighted summation of all basic causal contributions of each workstation in the PCBA table-mounted line to obtain the causal contribution of each workstation in the PCBA table-mounted line to the defect of the target PCB board; wherein, the workstation feature table is a correspondence table between workstations and workstation feature types in the PCBA table-mounted line.

[0018] Optionally, the cloud platform performs station-based source tracing of defects on the target PCB board based on the causal contribution of each station in the PCBA surface mount line to the defects. This includes: the cloud platform obtaining upstream causal contribution, midstream causal contribution, and downstream causal contribution based on the causal contribution of each station in the PCBA surface mount line to the defects; wherein, the upstream causal contribution includes the causal contribution of solder paste printing station and solder paste printing inspection station to the defects; the midstream causal contribution includes the causal contribution of placement station and placement position inspection station to the defects; and the downstream causal contribution includes the causal contribution of reflow oven and soldering quality inspection station to the defects; the cloud platform determines the root cause stage of the target PCB board based on the upstream, midstream, and downstream causal contribution, as well as a pre-set root cause stage judgment strategy, to perform station-based source tracing of defects on the target PCB board.

[0019] Secondly, embodiments of this application provide a defect tracing system for PCBA surface mount lines, comprising: a PCBA surface mount line and a cloud platform connected to each workstation in the PCBA surface mount line; the PCBA surface mount line is used to process target PCB boards, and when a defect type of the target PCB board is detected during the processing, it sends the defect type of the target PCB board to the cloud platform; the cloud platform, when receiving the defect type of the target PCB board fed back by the PCBA surface mount line, obtains cross-workstation feature data of the target PCB board during the PCBA surface mount line flow process based on a pre-set unique identifier of the target PCB board; and obtains a historical sample set based on the defect type of the target PCB board and a pre-built historical cross-workstation feature database; its In this process, the historical sample set includes historical cross-station feature data corresponding to multiple normal PCB boards and historical cross-station feature data corresponding to multiple defective PCB boards with the same defect type as the target PCB board. Based on the cross-station feature data of the target PCB board and the historical sample set, as well as a pre-set causal contribution statistical algorithm, the causal contribution of each station in the PCBA surface mounting line to the defect of the target PCB board is obtained. The causal contribution statistical algorithm is used to determine the causal contribution of each station in the PCBA surface mounting line to the defect of the target PCB board based on the distribution of the cross-station feature data of the target PCB board in the historical sample set. Based on the causal contribution of each station in the PCBA surface mounting line to the defect of the target PCB board, the defects of the target PCB board are traced back to their source at each station.

[0020] Optionally, the PCBA surface mount line includes: a solder paste printing station, a solder paste printing inspection station, a chip mounting station, a chip mounting position inspection station, a reflow oven, and a soldering quality inspection station arranged sequentially according to the process flow; wherein, the cloud platform is further used to generate station process data corresponding to each station in each PCBA surface mount line when receiving process parameters input by the user, and to send it to the corresponding station; the station process data includes solder paste printing process data, printing inspection process data, chip mounting station process data, chip mounting position inspection process data, reflow oven temperature control process data, and soldering quality inspection process data; the solder paste printing station is used to perform solder paste printing processing on the target PCB board according to the received solder paste printing process data; the solder paste printing inspection station is used to inspect the target PCB board according to the received printing inspection process data. The system checks whether the solder paste printing on the B-board is qualified and obtains the target solder paste printing inspection data; the placement station is used to perform placement processing on the target PCB board according to the received placement station process data; the placement position detection station is used to detect whether the placement position of the target PCB board is qualified according to the received placement position detection process data and obtains the placement position detection data; the reflow oven is used to perform soldering processing on the target PCB board according to the received reflow oven temperature control process data; the soldering quality inspection station is used to detect whether the soldering of the target PCB board is qualified according to the received soldering quality inspection process data and obtains the soldering quality inspection data, and when it is determined that the soldering of the target PCB board is unqualified, it obtains the defect type of the target PCB board based on the soldering quality inspection data, and uploads the defect type to the cloud platform.

[0021] Optionally, the PCBA surface mount line further includes: a laser marking machine disposed between the automatic board suction station and the solder paste printing station; the laser marking machine is used to mark a unique identifier at a preset position on the target PCB board; the solder paste printing station is also used to obtain the unique identifier of the target PCB board and map the solder paste printing process data to the unique identifier of the target PCB board; the solder paste printing inspection station is also used to obtain the unique identifier of the target PCB board and map the target solder paste printing inspection data to the unique identifier of the target PCB board; the placement station is also used to obtain the unique identifier of the target PCB board and map the placement station process data to the target solder paste printing inspection data. The unique identifier of the target PCB board is mapped; the chip placement detection station is also used to obtain the unique identifier of the target PCB board and map the chip placement detection process data to the unique identifier of the target PCB board; the reflow oven is also used to obtain the unique identifier of the target PCB board and map the reflow oven temperature control process data to the unique identifier of the target PCB board; the soldering quality inspection station is also used to obtain the unique identifier of the target PCB board and map the soldering quality inspection data to the unique identifier of the target PCB board; the cloud platform is also used to obtain the mapping processing results of each station in the PCBA surface mount line in real time.

[0022] (III) Beneficial Effects

[0023] This application provides a defect tracing method for PCBA surface mount lines. It utilizes the unique identifier of the target PCB board to accurately collect cross-station feature data across the entire process. Combined with defect type, it matches and constructs a historical sample set containing normal samples and similar defect samples from a historical cross-station feature database. Based on a causal contribution statistical algorithm and the distribution pattern of target PCB board features in the historical sample set, it quantifies the causal contribution of each production station to PCB board defects. This method fully considers the process coupling relationship between upstream and downstream stations, effectively avoids misjudgment bias caused by single-dimensional detection and judgment, and significantly improves the accuracy and objectivity of PCBA surface mount line defect tracing. It provides reliable data support for production line process rectification, parameter optimization, and defect source management. Attached Figure Description

[0024] Figure 1 This application provides a schematic flowchart of a method for tracing defects in PCBA surface mount lines.

[0025] Figure 2 This is a schematic diagram of a PCBA surface mount trace tracking interface provided in an embodiment of this application. Detailed Implementation

[0026] To better explain and facilitate understanding of this application, the following detailed description of the application is provided in conjunction with the accompanying drawings and specific embodiments.

[0027] PCBA surface mount production lines consist of multiple process stations sequentially, including solder paste printing, inspection, placement, reflow soldering, and post-reflow inspection. The process parameters at each station are easily affected by environmental fluctuations, equipment aging, and material differences, which can readily lead to various PCB board defects such as solder paste misalignment, placement misalignment, and cold solder joints. Current PCBA defect handling methods mostly involve simple post-defect reviews after detection, lacking cross-station correlation analysis capabilities and making it difficult to accurately pinpoint the true source station of the defect.

[0028] Currently, most conventional defect tracing methods rely solely on the detection data of a single workstation for independent identification, ignoring the process coupling and chain transmission effects between upstream and downstream workstations. This can easily lead to misjudging statistical correlations caused by process linkages as the root cause of defects, resulting in low tracing accuracy.

[0029] Therefore, the defect tracing method for PCBA surface mount lines provided in this application accurately collects cross-station feature data throughout the entire process using the unique identifier of the target PCB board. It also matches and constructs a historical sample set containing normal samples and similar defect samples from a historical cross-station feature database based on defect types. By relying on a causal contribution statistical algorithm combined with the distribution pattern of target PCB board features in the historical sample set, it quantifies the causal contribution of each production station to PCB board defects. This method fully considers the process coupling relationship between upstream and downstream stations, effectively avoids misjudgment bias caused by single-dimensional detection and judgment, and significantly improves the accuracy and objectivity of PCBA surface mount line defect tracing. It provides reliable data support for production line process rectification, parameter optimization, and defect source management.

[0030] To better understand the above technical solutions, exemplary embodiments of this application will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application can be understood more clearly and thoroughly, and that the scope of this application can be fully conveyed to those skilled in the art.

[0031] This application provides a defect tracing method for PCBA surface mount lines. The method is based on a pre-deployed PCBA surface mount line and a cloud platform. The PCBA surface mount line includes multiple workstations arranged sequentially according to the process flow. The method is as follows: Figure 1 As shown, it includes:

[0032] S1. When the cloud platform receives the defect type of the target PCB board from the PCBA surface mounting line, the cloud platform obtains the cross-workstation characteristic data of the target PCB board in the PCBA surface mounting line flow process based on the unique identifier pre-set in the target PCB board.

[0033] S2. The cloud platform obtains a historical sample set based on the defect type of the target PCB board and a pre-built historical cross-workstation feature database. The historical sample set includes historical cross-workstation feature data corresponding to multiple normal PCB boards and historical cross-workstation feature data corresponding to multiple defective PCB boards with the same defect type as the target PCB board.

[0034] S3, the cloud platform uses cross-station feature data and historical sample sets of the target PCB board, as well as a pre-set causal contribution statistical algorithm, to obtain the causal contribution of each station in the PCBA surface mounting line to the defects of the target PCB board; the causal contribution statistical algorithm is used to determine the causal contribution of each station in the PCBA surface mounting line to the defects of the target PCB board based on the distribution of cross-station feature data of the target PCB board in the historical sample set.

[0035] S4. The cloud platform traces the defects of the target PCB board by determining the causal contribution of each workstation in the PCBA surface mount line to the defects of the target PCB board.

[0036] This embodiment provides a defect tracing method for PCBA surface mount lines. It utilizes the unique identifier of the target PCB board to accurately collect cross-station feature data across the entire process. Combined with defect type, it matches and constructs a historical sample set containing normal samples and similar defect samples from a historical cross-station feature database. Based on a causal contribution statistical algorithm and the distribution pattern of target PCB board features in the historical sample set, it quantifies the causal contribution of each production station to PCB board defects. This method fully considers the process coupling relationship between upstream and downstream stations, effectively avoids misjudgment bias caused by single-dimensional detection and judgment, and significantly improves the accuracy and objectivity of PCBA surface mount line defect tracing. It provides reliable data support for production line process rectification, parameter optimization, and defect source management.

[0037] Optionally, in a specific embodiment, the PCBA surface mount line includes: a solder paste printing station, a solder paste printing inspection station, a chip mounting station, a chip mounting position inspection station, a reflow oven, and a soldering quality inspection station arranged sequentially according to the process flow;

[0038] When the cloud platform receives the defect type feedback from the PCBA surface mount line of the target PCB board, the cloud platform obtains the cross-workstation characteristic data of the target PCB board during the PCBA surface mount line flow based on the pre-set unique identifier of the target PCB board, including:

[0039] When the cloud platform receives the defect type of the target PCB board from the welding quality inspection station, it retrieves the corresponding solder paste printing process data, solder paste printing inspection data, SMT station process data, SMT position detection data, reflow oven temperature control process data, and welding quality inspection data based on the unique identifier pre-set for the target PCB board. The solder paste printing process data refers to the process data of the solder paste printing station during solder paste printing on the target PCB board; the solder paste printing inspection data refers to the inspection data from the solder paste printing inspection station; the SMT station process data refers to the process data of the SMT station during SMT assembly on the target PCB board; the SMT position detection data refers to the SMT position data detected by the SMT position detection station; the reflow oven temperature control process data refers to the process data of the reflow oven during welding on the target PCB board; and the welding quality inspection data refers to the welding quality data detected by the welding quality inspection station.

[0040] The cloud platform extracts features from solder paste printing process data, solder paste printing inspection data, chip placement station process data, chip placement position inspection data, reflow oven temperature control process data, and soldering quality inspection data. It then sorts the quantified features extracted from all data according to the process flow to obtain cross-station feature data of the target PCB board.

[0041] Specifically, in this embodiment, the PCBA surface mount line is set up with the following six stations in sequence according to the process flow: solder paste printing station → solder paste printing inspection station → chip mounting station → chip mounting position inspection station → reflow oven station → soldering quality inspection station.

[0042] The station includes several sections: a solder paste printing station equipped with a solder paste printer to automatically print solder paste onto the corresponding solder joints on the PCB board, preparing for component soldering; a solder paste printing inspection station equipped with an SPI solder paste printing inspection machine to inspect the quality of the solder paste printing, including thickness, area, and missed areas; a component placement station equipped with a component placement machine to accurately place the components to be placed onto the corresponding solder joints on the PCB board; a component placement position inspection station equipped with a pre-reflow AOI automatic optical inspection machine to inspect the accuracy of the component placement and whether there are any missed placements; a reflow oven station equipped with a reflow oven to solder and fix the placed electronic components onto the PCB board according to a preset temperature control profile; and a soldering quality inspection station equipped with a post-reflow AOI automatic optical inspection machine to inspect the soldering quality of the PCB boards coming out of the reflow oven and identify defects such as cold solder joints, bridging, tombstoning, and solder balls.

[0043] In addition, before the PCB board enters the solder paste printing station, there are automatic board feeding machines (to realize automatic board loading of bare boards) and laser marking machines (to mark a unique QR code at a selected position on the PCB board). These two devices provide a board-level unique identification basis for the collection of cross-station feature data.

[0044] Before the PCB board enters the first process station (solder paste printing station) of the PCBA surface mount line, a laser marking machine marks a unique identification QR code (unique identifier) ​​at a selected location on the PCB board. This QR code uses a high-fault-tolerant encoding format (such as Data Matrix ECC200) and contains a unique board-level number. The encoding rule is "production line number + date serial number + board serial number", for example, "SMT01-20260515-00001".

[0045] This unique identification QR code serves as the primary identifier for the target PCB board throughout the entire PCBA surface mount production process. It possesses the following characteristics: Uniqueness: Each PCB board has a globally unique QR code number, ensuring it won't be duplicated with any other PCB board and preventing confusion during subsequent data association; Durability: The QR code is permanently marked on the PCB board using laser engraving and will not be damaged or detached during the entire production process, such as solder paste printing or reflow soldering, ensuring accurate identification of the PCB board at every workstation; Readability: Each workstation's scanner can automatically read the QR code, with a reading speed meeting production line cycle time requirements (typically less than 0.5 seconds per board), thus avoiding becoming a production bottleneck.

[0046] When the cloud platform receives feedback from the welding quality inspection station regarding the defect type of the target PCB board, the cloud platform initiates the process of acquiring cross-station feature data. The specific steps are as follows:

[0047] After completing the soldering quality inspection of the target PCB board, the welding quality inspection station (post-reflow AOI) uploads the inspection results to the cloud platform via the production line collaborative control system (EAP). The inspection results include a unique identifier for the target PCB board and a defect type classification label. The defect type classification labeling rules are as follows:

[0048] Cold solder joint (defect type) - the solder joint does not form an effective intermetallic compound connection (judgment condition) - the solder joint looks normal but the electrical connection is unreliable (physical manifestation); Bridging (defect type) - an undesigned solder connection appears between adjacent solder joints (judgment condition) - short circuit (physical manifestation); Tombstone (judgment condition) - one end of the chip component is lifted and detached from the pad (judgment condition) - the component stands upright (physical manifestation); Solder ball (judgment condition) - isolated solder balls appear around the solder joint (judgment condition) - potential short circuit risk (physical manifestation), etc.

[0049] After receiving the defect type trigger signal, the cloud platform uses the unique identifier of the target PCB board as the primary key to initiate the cross-workstation feature data retrieval and assembly process.

[0050] Based on the unique identifier of the target PCB board, the cloud platform retrieves the corresponding solder paste printing process data from the data interface of the solder paste printing station. This solder paste printing process data reflects the operating parameters and process execution status of the solder paste printer during the solder paste printing process on the target PCB board. Specifically, this includes, but is not limited to, squeegee pressure, printing speed, demolding speed, and cleaning cycle. This data is automatically recorded by the solder paste printer during the printing process and uploaded in real-time to the production line collaborative control system via Modbus / TCP or HTTP / Socket interfaces, and stored in association with the unique identifier of the PCB board.

[0051] Based on the unique identifier of the target PCB board, the cloud platform retrieves the corresponding solder paste printing inspection data from the data interface of the Solder Paste Printing Inspection Station (SPI). This solder paste printing inspection data is the output data from the SPI after performing solder paste printing quality inspection on the target PCB board, reflecting the actual quality status of the solder paste printing. Specifically, it includes, but is not limited to: mean solder paste thickness, solder paste thickness variance, solder paste area deviation rate, and missed solder paste rate. This data is acquired by the SPI device through 3D laser scanning, uploaded via a CSV file or MySQL relational database interface, and stored in association with the unique identifier of the PCB board.

[0052] Based on the unique identifier of the target PCB board, the cloud platform retrieves the corresponding SMT process data from the data interface of the SMT station.

[0053] The pick-and-place station process data refers to the process data generated when the pick-and-place station performs component placement on the target PCB board. It reflects the equipment operating parameters and placement execution status of the pick-and-place machine during the placement process. Specifically, this includes, but is not limited to: placement offset X, placement offset Y, placement angle deviation, and missing placement rate. This data is automatically recorded by the pick-and-place machine during the placement process, uploaded via an HTTP / RESTful interface, and stored in association with a unique identifier on the PCB board.

[0054] Based on the unique identifier of the target PCB board, the cloud platform retrieves the corresponding component placement detection data from the data interface of the component placement detection station (AOI before reflow). This placement data, detected by the AOI station, reflects the component placement quality after placement. Specifically, it includes, but is not limited to: placement offset pass rate, component missing rate, and polarity reversal rate. This data is acquired by the AOI equipment through multi-angle light source imaging and intelligent algorithm analysis, uploaded via HTTP / Socket interface, and stored in association with the PCB board's unique identifier.

[0055] Based on the unique identifier of the target PCB board, the cloud platform retrieves the corresponding reflow oven temperature control process data from the reflow oven's data interface. This reflow oven temperature control process data reflects the actual temperature profile experienced by the target PCB board during soldering in the reflow oven. This includes, but is not limited to, peak temperature deviation, heating slope deviation, holding zone temperature deviation, and cooling slope deviation. This data is collected in real-time by the reflow oven temperature control system via in-oven thermocouples, uploaded through a Modbus / TCP interface, and stored in association with the PCB board's unique identifier.

[0056] Based on the unique identifier of the target PCB board, the cloud platform retrieves the corresponding soldering quality inspection data from the data interface of the soldering quality inspection station (post-reflow AOI). The soldering quality inspection data reflects the final soldering quality status of the target PCB board, as detected by the soldering quality inspection station. This includes, but is not limited to, cold solder joint rate, bridging rate, tombstoning rate, and solder ball rate. This data is acquired by the post-reflow AOI equipment through multi-angle light source imaging and intelligent algorithm analysis, uploaded via HTTP / RESTful interface, and stored in association with the unique identifier of the PCB board.

[0057] Each of the above features is a workstation feature type proposed in this embodiment.

[0058] The cloud platform extracts features from the raw data obtained from the six workstations, transforming the original equipment operation records and test results into standardized quantitative features. Feature extraction follows these principles: board-level aggregation: raw data is measured at test points or components (e.g., the thickness measurement of each pad by SPI, the offset of each component by the pick-and-place machine), and needs to be aggregated into board-level statistical features (mean, variance, pass rate, etc.), with cross-workstation correlation performed on a board-by-board basis; dimensionless unification: the dimensions of data from each workstation are different (μm, mm, °C, %), and the original dimensions are retained after feature extraction. Dimensional differences are eliminated during subsequent mutual information calculation through equal-frequency binning discretization; information integrity: feature extraction considers both central tendency (mean) and dispersion (variance, pass rate), ensuring that it reflects both the overall state of the process and the fluctuations in process stability.

[0059] Subsequently, the cloud platform extracts all the quantified features from the data features, arranges them sequentially according to the process flow of the PCBA surface mounting lines, and assembles them into cross-station feature data for the target PCB board.

[0060] The structure of cross-station feature data is: x board =[x 印刷 ,x SPI ,x 贴片 ,x 炉前AOI ,x 回流炉 x 炉后AOI ]; Expanded to: x board =[Squeegee pressure, printing speed, demolding speed, cleaning cycle, average solder paste thickness, solder paste thickness variance, solder paste area deviation rate, missed brush rate, component offset X, component offset Y, component angle deviation, missed placement rate, placement offset pass rate, component missing rate, polarity inversion rate, peak temperature deviation, heating slope deviation, insulation zone temperature deviation, cooling slope deviation, cold solder joint rate, bridging rate, tombstoning rate, solder ball rate].

[0061] The cross-station feature data consists of 23 dimensions, each corresponding to a station feature type, i.e., a specific process or testing quantitative indicator.

[0062] The devices at each workstation on the PCBA surface mount production line come from different manufacturers and use different communication protocols and data formats. The cloud platform needs to modify the interfaces of all types of devices to uniformly adapt them to a standard interface format, such as the unified data summary interface format of EAP.

[0063] Because the sampling frequencies of different workstations are different (e.g., solder paste printers record once per board, reflow oven temperature control systems record once per second, and SPI takes about 60 seconds to detect per board), the directly acquired raw data is inconsistent in terms of time granularity. The cloud platform achieves time window alignment in the following ways:

[0064] Based on the PCB board's transit time (entry and exit times) on each equipment, all raw data for that board on each equipment is aggregated into single-board-level statistics. For example, a reflow oven generates 60 temperature sampling points during its 60-second transit time, and the average value is taken as the board's peak temperature deviation; SPI scans all pads on the entire board within 60 seconds, and the mean and variance of all pad thickness measurements are taken as the mean and variance of the board's solder paste thickness. This aggregation method based on the PCB board's transit time ensures that each board has only one corresponding set of feature values ​​at each workstation, eliminating the impact of sampling frequency differences on cross-workstation data correlation.

[0065] After acquiring data from each workstation, the cloud platform verifies the completeness and reasonableness of the data:

[0066] Integrity verification: Check whether the data of the target PCB board is complete across all six workstations. If data is missing at a certain workstation (e.g., due to equipment communication interruption causing data failure to upload), the board is marked as "incomplete data" and will not participate in subsequent causal contribution calculations to avoid statistical bias caused by missing data.

[0067] Reasonableness verification: Check whether each feature value is within the physically reasonable range (e.g., the average solder paste thickness should be between 50 and 300 μm, and the missed brush rate should be between 0 and 100%). If a feature value exceeds the reasonable range, it is marked as an outlier and replaced with the average value of that feature on the most recent N normal boards to ensure the availability of feature data across workstations.

[0068] The embodiments provided in this application standardize feature extraction, dimensionless regularization, and multidimensional quantification of the original data of each workstation according to the process flow, and orderly assemble the features to form 23-dimensional cross-workstation feature data with standardized structure, unified dimensions, and complete information. This not only fully preserves the process status and quality fluctuation information of each process, but also realizes the standardized fusion of heterogeneous data from different workstations. This lays the foundation for subsequent statistical calculation of causal contribution based on historical sample sets, accurate quantification of the impact of each workstation on defects, and high-precision defect tracing.

[0069] Optionally, in a specific embodiment, the cross-station feature data of the target PCB board and each historical cross-station feature data include multiple station feature types;

[0070] The cloud platform obtains historical sample sets based on the defect types of the target PCB board and a pre-built historical cross-workstation feature database, including:

[0071] Based on the defect type of the target PCB board, the cloud platform filters out the historical cross-station feature data corresponding to all defective PCB boards with the same defect type as the target PCB board from the pre-built historical cross-station feature database, as well as the historical cross-station feature data corresponding to all normal PCB boards without defects; and marks all defective PCB boards with defect labels.

[0072] The cloud platform uses a pre-set equal-frequency binning discretization algorithm to discretize each selected historical cross-workstation feature data. The equal-frequency binning discretization strategy includes: for each workstation feature type, statistically analyzing the feature values ​​corresponding to that workstation feature type in all historical cross-workstation feature data to obtain the feature value distribution corresponding to each workstation feature type; and based on the feature value distribution corresponding to each workstation feature type, dividing each workstation feature type into a preset number of discrete intervals so that the feature values ​​corresponding to each workstation feature type in all historical cross-workstation feature data are mapped to the discrete intervals.

[0073] A historical sample set is constructed based on all historical cross-station feature data after discrete processing and all defect labels.

[0074] Before performing defect root cause analysis, the cloud platform needs to pre-build a historical cross-workstation feature database as the statistical basis for subsequent causal contribution calculations. The process of building the historical cross-workstation feature database is as follows:

[0075] During the daily production operation of the PCBA surface mount line, the cloud platform continuously collects process data and inspection data for each PCB board at each workstation. Following the cross-workstation feature data acquisition method described in the previous embodiment, it assembles cross-workstation feature data for each PCB board and stores the board's unique identifier, cross-workstation feature data, and soldering quality inspection results together in the historical cross-workstation feature database. The data structure for each historical record is [PCB ID (unique identifier), x1~x 23 (Cross-station feature data), is defect (whether it is defective, defective PCB boards are marked as 1, normal PCB boards are marked as 0), defect type (defect type, normal boards are marked as 0)).

[0076] As production continues, records in the historical cross-station feature database accumulate. In this embodiment, the production data of the most recent 30 days is used as the statistical window, and the database stores a total of 8,580 historical records.

[0077] Once the cloud platform receives the defect type of the target PCB board, it uses that defect type as the search criterion to filter out historical data of the same type from the historical cross-workstation feature database and construct the sample set required for defect analysis.

[0078] The cloud platform filters all historical records from the historical cross-workstation feature database where the "is defect" field is 1 and the "defect type" field is the same as the defect type of the target PCB board. It then extracts the historical cross-workstation feature data corresponding to these records as samples of defective PCB boards.

[0079] The defect type matching rules are as follows: If the target PCB board's defect type is cold solder joint, then all cold solder joint defect records in the historical database are filtered; if the target PCB board's defect type is bridging, then all bridging defect records in the historical database are filtered; if the target PCB board's defect type is tombstoning, then all tombstoning defect records in the historical database are filtered; if the target PCB board's defect type is solder ball, then all solder ball defect records in the historical database are filtered. Since the same PCB board may have multiple defect types simultaneously (e.g., a board with both cold solder joint and solder ball defects is detected), for such multi-defect boards, as long as it contains the same defect type as the target PCB board, it is included in the filtering range.

[0080] The cloud platform filters historical records from the cross-station feature database, selecting all records where the "is defect" field is 0. It then extracts the corresponding historical cross-station feature data for these records, using them as samples of normal PCB boards. A normal PCB board is defined as one that did not detect any defects at the welding quality inspection station. The cloud platform assigns a defect label D=1 to each selected defective PCB board and a defect label D=0 to each normal PCB board. The defect label is a binary variable used to distinguish between defective and normal samples in subsequent mutual information calculations.

[0081] In historical cross-station feature data, the feature values ​​of each station feature type are continuous variables, with values ​​ranging from continuous intervals, making them unsuitable for direct probability statistics. For example, the average solder paste thickness ranges from 105 to 138 μm, theoretically having an infinite number of possible values. However, directly calculating the frequency of each precise value would result in most values ​​appearing only once or even zero times, failing to provide an effective probability estimate. Therefore, it is necessary to discretize the continuous feature values ​​of each station feature type, mapping the continuous values ​​to a finite number of discrete intervals. This ensures that each interval contains a sufficient number of samples to support stable probability estimation.

[0082] The core principle of equal-frequency binning discretization is that each discrete interval contains approximately the same number of samples. Compared to equal-width binning (where each interval has an equal range of values), equal-frequency binning has the following advantages:

[0083] Avoid empty or sparse bins: When the distribution of feature values ​​is skewed (e.g., the missing rate is mostly 0 and a few are positive), equal-width binning will result in no or very few samples in some intervals, making the probability estimation unstable; equal-frequency binning ensures that the number of samples in each interval is similar, making the probability estimation more reliable.

[0084] Adaptive feature distribution: The interval boundaries of equal-frequency binning are automatically determined by the data distribution. The intervals are narrower in densely populated areas and wider in sparsely populated areas, making full use of the information content of the data.

[0085] A unified probability benchmark: Since the number of samples in each interval is roughly equal, the prior probability (marginal probability) of each interval is close to 1 / K, making the contribution weight of each interval to the result more balanced in the mutual information calculation.

[0086] For each of the 23 workstation feature types, the cloud platform independently performs the following steps:

[0087] The cloud platform iterates through all 8580 historical records, extracting feature values ​​for each workstation's characteristic type from each record to form a feature value set corresponding to that workstation's characteristic type. This feature value set is then sorted in ascending order to obtain the feature value distribution for that workstation's characteristic type. The feature value distribution reflects the range, central tendency, and dispersion of the workstation's characteristic in historical production data. For example, the feature value distribution of the mean solder paste thickness may exhibit an approximately normal distribution, with dense samples near the mean and sparse samples at both ends; the feature value distribution of the missing solder paste rate may exhibit a severely right-skewed distribution, with the vast majority of boards having a missing solder paste rate of 0, and a few boards having a positive missing solder paste rate.

[0088] The cloud platform presets the number of bins K (the number of discrete intervals), and in this embodiment, K=5. The sorted feature value set is divided into K groups, each containing N / K samples (where N is the total number of selected samples, which is 8580 in this example, and N / K=1716). Specifically, the feature value of the i×(N / K)th sample after sorting is taken as the boundary value between the i-th interval and the (i+1)-th interval (i=1,2,...,K-1), generating a total of K-1 boundary values, thus dividing the feature value range into K discrete intervals. For each feature value of this workstation feature type in a historical record, it is determined which discrete interval it falls into, and the feature value is replaced with the corresponding interval number (1~K). After mapping, the workstation feature type is transformed from a continuous variable into a discrete variable with values ​​from 1 to K.

[0089] Taking the mean solder paste thickness as an example (which approximates the characteristics of a normal distribution), the quantiles of the mean solder paste thickness in 8580 records are as follows:

[0090] Quantitation 0% has a characteristic value of 105.0 μm; quantitation 20% has a characteristic value of 114.2 μm; quantitation 40% has a characteristic value of 118.6 μm; quantitation 60% has a characteristic value of 122.1 μm; quantitation 80% has a characteristic value of 126.8 μm; and quantitation 100% has a characteristic value of 138.0 μm.

[0091] The five discrete intervals are: [105.0, 114.2), with 1716 samples; [114.2, 118.6), with 1716 samples; [118.6, 122.1), with 1716 samples; [122.1, 126.8), with 1716 samples; and [126.8, 138.0), with 1716 samples.

[0092] The average solder paste thickness of the target PCB board is 118.5μm, which falls within interval 2. After mapping, x5=2.

[0093] Another example is the missed labeling rate (a characteristic of a severely right-skewed distribution). Among 8580 records, the ranking quantiles of the missed labeling rate are as follows: quantile 0%, characteristic value 0%; quantile 20%, characteristic value 0%; quantile 40%, characteristic value 0%; quantile 60%, characteristic value 0%; quantile 80%, characteristic value 0.2%; quantile 100%, characteristic value 5.0%.

[0094] Since the omission rate for a large number of samples is 0%, the 20th, 40th, and 60th percentile values ​​are all 0. During equal-frequency binning, samples with the same value are assigned to the same interval, resulting in an uneven number of samples per interval. The cloud platform's strategy for handling identical values ​​is to assign all samples with the same value to the same interval and readjust the boundaries between adjacent intervals to balance the number of samples as much as possible. The adjusted five discrete intervals are: [0%, 0%], 5136 samples; (0%, 0.1%], 1716 samples; (0.1%, 0.3%], 1020 samples; (0.3%, 0.8%], 468 samples; (0.8%, 5.0%], 240 samples. Due to the severely right-skewed distribution of the missing component rate, the number of samples in interval 1 (no missing components) is much larger than in the other intervals, and the number of samples in the five intervals is not completely equal. This is an inherent limitation of equal-frequency binning when facing a large number of repeated values, but compared to equal-width binning (which may result in almost no samples in the later intervals), equal-frequency binning can still ensure that there are enough samples in each interval to support the probability estimation. The missing component rate of the target PCB board is 0%, falling into interval 1 (no missing components), and after mapping, x 12 =1.

[0095] After equal-frequency binning discretization of all 23 workstation feature types, the 23-dimensional continuous feature values ​​of each historical record are replaced with the corresponding discrete interval numbers (1~5). The cloud platform combines the 23-dimensional discretized feature values ​​of each record with the defect label to form a sample record: sample record = [ , ,…, ,D]. Among them, Let D be the discrete interval number of the feature type of the i-th workstation, and D be the defect label (1=defect, 0=normal). The cloud platform will aggregate all 8580 discretized sample records to construct a historical sample set. The historical sample set is an 8580-row × 24-column matrix (23-dimensional discrete features + 1-dimensional defect label), with each row corresponding to a historical PCB board.

[0096] It is important to note that before performing equal-frequency binning, the cloud platform first performs outlier detection on the feature values ​​of each workstation's feature type. Based on the 3σ criterion, it identifies and removes extreme values ​​that significantly deviate from the normal range to avoid adverse effects of extreme values ​​on the binning boundaries (e.g., an extremely large value causing the last interval to contain only one sample). The removed outliers are not included in the calculation of the binning boundaries, but they are still mapped to the nearest interval after the binning boundaries are determined.

[0097] The cross-workstation feature data of the target PCB board itself also needs to be discretized, but the target PCB board does not participate in the calculation of the binning boundaries. The binning boundaries are determined entirely by the historical sample set, and the feature values ​​of the target PCB board are mapped to the corresponding discrete intervals based on the binning boundaries determined by the historical sample set. This ensures that the target board and historical samples are comparable under the same discretization standard. The binning boundary values ​​of each workstation feature type are stored in the cloud platform for the discretization mapping of subsequently added PCB board data. When the cumulative amount of data in the historical cross-workstation feature database exceeds a preset threshold (e.g., the added data exceeds 20% of the original data volume), or when the time since the last binning boundary calculation exceeds a preset time (e.g., 30 days), the cloud platform automatically recalculates the binning boundaries to ensure that the boundary values ​​reflect the latest production data distribution.

[0098] This embodiment achieves structured separation of defective and normal PCBs by filtering PCBs with the same type of defects from a historical cross-station feature database based on defect type and labeling them with defect tags, thus providing a statistical comparison basis for subsequent mutual information calculation.

[0099] Optionally, in a specific embodiment, the cloud platform, based on cross-workstation feature data and historical sample sets of the target PCB board, and a pre-set causal contribution statistical algorithm, obtains the causal contribution of each workstation in the PCBA surface mounting line to the defects of the target PCB board, including:

[0100] The cloud platform maps the feature values ​​of each workstation feature type in the cross-workstation feature data of the target PCB board to the corresponding discrete interval;

[0101] Based on all historical cross-station feature data and all defect type labels in the historical sample set, the cloud platform obtains the joint probability distribution, marginal distribution of station features, and marginal distribution of defect labels for each feature type in the historical sample set. Specifically, the joint probability distribution represents the probability distribution of feature values ​​within each discrete interval for historical cross-station feature data corresponding to a particular station feature type, where the feature values ​​are either with or without defect labels. The marginal distribution of station features represents the probability distribution of feature values ​​within each discrete interval for historical cross-station feature data corresponding to a particular station feature type. The marginal distribution of defect labels represents the probability distribution of whether or not a feature data has a defect label within all historical cross-station feature data corresponding to a particular station feature type.

[0102] The cloud platform obtains mutual information corresponding to each workstation feature type based on the joint probability distribution, normal joint probability distribution, workstation feature marginal distribution, and defect label marginal distribution corresponding to each workstation feature type in the historical sample set.

[0103] The cloud platform obtains the causal contribution of each station in the PCBA surface mounting line to the defects of the target PCB board based on the cross-station feature data of the target PCB board and the mutual information corresponding to the feature type of each station.

[0104] Furthermore, based on the joint probability distribution, normal joint probability distribution, marginal distribution of workstation features, and marginal distribution of defect labels corresponding to each workstation feature type in the historical sample set, the cloud platform obtains the mutual information corresponding to each workstation feature type, including:

[0105] The cloud platform obtains the mutual information between each workstation feature type and defect type based on the joint probability distribution, marginal distribution of workstation features, and marginal distribution of defect labels corresponding to each workstation feature type in the historical sample set, as well as the pre-set Formula 1; Formula 1 is:

[0106] ;

[0107] Where r(X) i Let p(k,d) be the mutual information between the workstation feature type and the defect type at index i, k be the index of the discrete interval, M be the number of discrete intervals, and d be the index of whether or not a defect label is present. p(k) is the probability that the feature value in the marginal distribution of the workstation feature type falls within the k-th discrete interval, p(d) is the probability that the feature value in the marginal distribution of the defect label of the workstation feature type has a defect label / does not have a defect label, and p(k,d) is the probability that the feature value in the joint probability distribution of the workstation feature type has / does not have a defect label and is within the k-th discrete interval.

[0108] Furthermore, based on the cross-station feature data of the target PCB board and the mutual information corresponding to the feature type of each station, the cloud platform obtains the causal contribution of each station in the PCBA surface mounting line to the defects of the target PCB board, including:

[0109] The cloud platform normalizes the mutual information corresponding to all workstation feature types to obtain the basic causal contribution of each workstation feature type; the sum of the basic causal contribution of all workstation feature types is 1.

[0110] The cloud platform obtains the deviation weight corresponding to each workstation feature type based on the cross-workstation feature data corresponding to the target PCB board and the historical cross-workstation feature data corresponding to all normal PCB boards in the historical sample set, as well as the pre-set Formula 2; Formula 2 is:

[0111] ;

[0112] Among them, w i x is the deviation weight corresponding to the feature type of the workstation with index i. i,目标板 Let i be the feature value of the cross-station feature data corresponding to the target PCB board at the station feature type with index i. This represents the mean of the feature values ​​for all normal PCB boards at the workstation feature type with index i. The standard deviation of the eigenvalues ​​for all normal boards on the feature type of the workstation at index i;

[0113] The cloud platform uses the deviation weight corresponding to the feature type of each workstation and the pre-set workstation feature table to perform a weighted sum of all the basic causal contributions of each workstation in the PCBA table-mounting line, so as to obtain the causal contribution of each workstation in the PCBA table-mounting line to the defects of the target PCB board. The workstation feature table is a correspondence table between workstations and workstation feature types in the PCBA table-mounting line.

[0114] Specifically, after acquiring the cross-workstation feature data of the target PCB board, the cloud platform maps the continuous feature values ​​of each workstation feature type to the corresponding discrete interval determined when dividing the historical sample set into equal-frequency bins.

[0115] The cloud platform obtains three probability distributions corresponding to each workstation feature type based on all historical cross-workstation feature data and all defect labels in the historical sample set. The joint probability distribution is the probability distribution of the feature values ​​corresponding to the historical cross-workstation feature data with or without defect labels within each discrete interval in all historical cross-workstation feature data corresponding to the workstation feature type.

[0116] For the feature type of workstation with index i, the joint probability distribution contains two sets of probability values:

[0117] The joint probability distribution of defects is given by the probability p(X) of the feature value of the workstation's feature type falling into the k-th discrete interval in historical samples with defect labels (D=1). i =k,D=1)=N(X i =k,D=1) / N total , where N(X) i =k,D=1) represents the number of samples in the historical sample set whose feature type value of this workstation falls into the k-th interval and is marked as defect, and N total Let be the number of samples in the historical sample set; in the normal joint probability distribution, among the historical samples without defect labels (D=0), the probability p(X) of the feature value of this workstation feature type falling into the k-th discrete interval. i =k,D=0)=N(X i=k,D=0) / N total , where N(X) i =k, D=0) represents the number of samples in the historical sample set whose feature type value falls into the k-th interval and is marked as normal. The joint probability distribution reflects the actual frequency of the simultaneous occurrence of each discrete interval of the workstation feature with the defect / normal state. When the joint probability of defect in a certain interval is significantly higher than the joint probability of normal in that interval (relative to the expected marginal probability of each), it indicates that there is a positive statistical association between that interval and the defect.

[0118] The marginal distribution of workstation features is the probability distribution of the feature values ​​of all historical cross-workstation feature data corresponding to the workstation feature type within each discrete interval, without considering the value of the defect label: p(X i =k)=N(X i =k) / N total , where N(X) i =k) ​​represents the total number of samples in the historical sample set whose feature type value falls within the k-th interval, including defective and normal samples. Due to the use of equal-frequency binning, the number of samples in each interval is approximately equal, and the marginal distribution of the station's features is approximately uniform. However, in reality, due to the allocation of identical values ​​and boundary adjustments, the marginal probabilities of each interval may have slight differences; therefore, the calculation is still based on the actual count.

[0119] The marginal distribution of defect labels is the probability distribution of having or not having defect labels among all historical cross-station feature data corresponding to the station feature type: p(D=1)=N(D=1) / N total p(D=0)=N(D=0) / N total Where N(D=1) is the total number of defective samples in the historical sample set, and N(D=0) is the total number of normal samples. The marginal distribution of defect labels reflects the basic probability of defects occurring in the historical sample set and is independent of any specific workstation feature type.

[0120] The cloud platform obtains the mutual information between each workstation feature type and defect type based on the joint probability distribution, marginal distribution of workstation features, and marginal distribution of defect labels corresponding to each workstation feature type, as well as the pre-set Formula 1. Formula 1 is:

[0121] .

[0122] Its physical meaning is to traverse all discrete intervals k of the workstation feature type and all values ​​d of the defect label, and for each group (k,d), calculate the ratio of the joint probability p(k,d) to the product of the two marginal probabilities p(k)×p(d), take the logarithm to the base 2 and then weight it with the joint probability p(k,d), and sum the weighted results for all combinations.

[0123] Taking the workstation feature type x5 "average solder paste thickness" as an example, the probability distribution is substituted into Formula 1 and calculated term by term: k=1, d=1: p(1,1)=0.0056, p(1)×p(1,label)=0.003636, ratio=0.00956 / 0.003636=2.629, log2(2.629)=1.395, term value=0.00956×1.395=0.01334; Similarly, k=1, d=0, term value -0.00836; k=2, d=1, term value 0.00126; k=2, d=0, term value -0.00113; k=3, d=1, term value -0.00139; k=3, d=0, term value 0.00174; k=4, d=1, term value -0.00193; k=4, d=1, term value 0.00343; k=5, d=1, term value -0.00139; k=5, d=0, term value 0.00462. Summation:

[0124] r(x5,d)=0.01334-0.00836+0.00126-0.00113-0.00139+0.00174-0.00193+0.00343-0.00139=0.01019bit.

[0125] The above calculations were performed on all 19 upstream station feature types (excluding the 4 features of the welding quality inspection station, as they are defects themselves rather than upstream process parameters) to obtain the mutual information value of each station feature type.

[0126] The cloud platform uses cross-workstation feature data of the target PCB board and historical cross-workstation feature data of all normal PCB boards in the historical sample set, along with a pre-set formula (Formula 2), to obtain the deviation weight corresponding to each workstation feature type. Formula 2 is:

[0127] .

[0128] The deviation weight is essentially the absolute value of the Z-score, reflecting the degree to which the target PCB board deviates from the mean of normal boards in this workstation feature, measured in the standard deviation of normal boards. The larger the deviation weight, the more abnormal the target PCB board is in this feature, and the more significant the causal contribution of this feature to the current defect of the target PCB board is; a deviation weight of zero indicates that the target board is consistent with the mean of normal boards in this feature, and this feature has no individual contribution to the current board defect.

[0129] The cloud platform normalizes the mutual information values ​​corresponding to all workstation feature types to obtain the basic causal contribution for each workstation feature type. The cloud platform then multiplies the basic causal contribution for each workstation feature type by its corresponding deviation weight to obtain the personalized causal contribution for each workstation feature type. Based on a pre-set workstation feature table, the cloud platform sums the personalized causal contribution of all workstation feature types within the same workstation to obtain the causal contribution of each workstation in the PCBA table-mounted wiring to the target PCB board defect.

[0130] For example, if a solder paste printing station has squeegee pressure, printing speed, demolding speed, and cleaning cycle, then the individual causal contribution values ​​of x1 to x4 are added together.

[0131] This embodiment quantifies the statistical correlation strength between various process features and defects, enabling automatic identification of the process factors with the greatest impact on defects from a large amount of historical data without relying on manual experience. Instead, it calculates deviation weights based on the deviation between the feature values ​​of the target PCB board and the statistical features of normal boards, combining historical statistical patterns with the individual characteristics of the current target board. This allows the causal contribution to reflect both statistical causal relationships at the group level and personalized adjustments for the specific abnormal state of each board. By summarizing the feature-level causal contribution into the workstation-level causal contribution through the workstation feature table, it achieves a mapping from fine-grained features to coarse-grained workstations, enabling the traceability results to directly correspond to operable and maintainable physical workstations on the production line, providing clear direction for operation and maintenance decisions.

[0132] Furthermore, in the PCBA surface mount line, each workstation is set up sequentially according to the process flow. Process deviations at upstream workstations can indirectly affect downstream defects through midstream workstations along the process flow. Therefore, the causal contribution of each workstation's characteristics to defects calculated by the cloud platform based on the mutual information formula may contain spurious causal relationships. Therefore, it is necessary to eliminate spurious causal relationships.

[0133] First, based on the process flow of PCBA surface mount lines, the cloud platform determines that the causal transmission direction between features is unidirectional: upstream station features can only indirectly affect downstream defects through midstream station features, and midstream station features can only indirectly affect defects through downstream station features. The cloud platform pre-constructs a set of candidate indirect transmission paths based on process knowledge. Each candidate indirect transmission path consists of one upstream station feature and one midstream station feature, denoted as (X... m →X n →D), representing the upstream station characteristics X m The impact of defect D may be transmitted through intermediate station feature X. n Indirect transmission. The cloud platform is based on features X from historical sample sets. m The feature X is calculated according to Formula 1, based on the sample distribution and defect label distribution within each discrete interval. mMutual information r(X) between the defect label D and the defect label D m ;D). This value has already been obtained in the causal contribution statistics algorithm and can be called directly. Similarly, calculate r(X). n ;D); Calculate the mutual information r(X) of the joint distribution of upstream station features and intermediate station features on the defect label. m ,X n ;D), the calculation formula is:

[0134] ;

[0135] Where, k m For feature X m Discrete interval index, k n For feature X n The discrete interval index, where d is the value of the defect label.

[0136] Based on the above calculation results, the cloud platform calculates the degree of independent association between the features of the upstream workstation and the defect, i.e., r(X), using the feature-level conditional mutual information algorithm. m ,X n ;D)-r(X n ;D), meaning the amount of additional information that upstream station features can still provide regarding the occurrence of a defect, given the known state of intermediate station features. A larger value indicates a stronger independent and direct influence of the feature on the defect; a value close to zero indicates that the influence of the feature on the defect is entirely transmitted indirectly through the feature. The cloud platform compares the feature-level conditional mutual information with a preset feature-level indirect association threshold to determine whether the candidate path has a spurious causal association. The feature-level indirect association threshold is determined based on the product of the unconditional mutual information of the upstream station features and a preset proportional coefficient. When it is less than the threshold, it is determined that the influence of the upstream station features on the defect is entirely transmitted indirectly through intermediate station features along the path, and the causal contribution of the upstream station features on this path is a spurious causal association. The cloud platform transfers the basic causal contribution of the upstream station features to the intermediate station features and sets the adjusted basic causal contribution of the features to zero. When it is greater than or equal to the threshold, it is determined that the upstream station features have an independent and direct contribution to the defect on this path, and it is not a spurious causal association; the basic causal contribution of the upstream features remains unchanged. However, the cloud platform further decomposes the basic causal contribution of upstream workstation features into direct contribution and indirect contribution components. Specifically, the direct contribution percentage = feature-level conditional mutual information / r(X) m ;D), indirect contribution ratio = [r(X m ;D)-Feature-level conditional mutual information] / r(X m(D) The cloud platform deducts the indirect contribution of upstream station features from their basic causal contribution and adds it to the basic causal contribution of intermediate station features. After performing feature-level conditional mutual information tests on all candidate indirect transmission paths, the cloud platform deducts the indirect contribution of each feature from the original feature and transfers it to the corresponding intermediate feature to obtain the adjusted basic causal contribution of each feature.

[0137] Optionally, in one specific embodiment, the cloud platform performs workstation-based source tracing of defects on the target PCB board based on the causal contribution of each workstation in the PCBA surface mount line to the defects of the target PCB board, including:

[0138] The cloud platform obtains upstream causal contribution, midstream causal contribution, and downstream causal contribution based on the causal contribution of each station in the PCBA surface mount line to the defects of the target PCB board. Among them, the upstream causal contribution includes the causal contribution of the solder paste printing station and the solder paste printing inspection station to the defects of the target PCB board; the midstream causal contribution includes the causal contribution of the placement station and the placement position inspection station to the defects of the target PCB board; and the downstream causal contribution includes the causal contribution of the reflow oven and the soldering quality inspection station to the defects of the target PCB board.

[0139] The cloud platform determines the root cause stage of the target PCB board based on upstream, midstream, and downstream causal contribution rates, as well as a pre-set root cause stage judgment strategy, in order to trace the defects of the target PCB board at the workstation.

[0140] Specifically, based on a pre-set workstation feature table, the cloud platform summarizes the causal contribution of each workstation in the PCBA surface mounting line to defects on the target PCB board, dividing it into three stages according to the process flow. The workstation feature table defines the correspondence between workstations and their feature types in the PCBA surface mounting line, as well as the relationship between workstations and stages. The cloud platform sums the causal contributions of all workstations within the same stage to obtain the causal contribution for each stage.

[0141] The cloud platform determines the root cause stage of the target PCB board based on upstream, midstream, and downstream causal contributions, as well as a pre-set root cause stage judgment strategy. Specifically, the root cause stage judgment strategy comprises three cascaded judgment rules, executed sequentially from highest to lowest priority. Once a rule at a certain level is met, the root cause stage is determined.

[0142] First, the cloud platform determines whether any of the three stages has a causal contribution greater than or equal to a preset dominant threshold (ranging from 0.45 to 0.55). If so, that stage is identified as the root cause stage of the target PCB board. Second, when the causal contribution of all three stages is less than the dominant threshold, the cloud platform sorts the causal contribution of the three stages from largest to smallest, obtaining a first candidate stage, a second candidate stage, and a third candidate stage. The cloud platform calculates the contribution ratio of the first candidate stage to the second candidate stage and compares the result with a resolution threshold (ranging from 1.3 to 1.7). If the ratio is greater than the first candidate stage, the first candidate stage is determined to have a significant advantage over the second candidate stage, and the first candidate stage is identified as the root cause stage. If the ratio is less than the first candidate stage, the contributions of the first candidate stage and the second candidate stage are determined to be close, and both have a significant causal effect on the occurrence of the defect. The first candidate stage and the second candidate stage are identified as the joint root cause stage, forming a two-stage joint root cause determination result. Finally, when the causal contribution of each of the three stages is less than the dominant threshold, and the difference between the causal contribution of the third candidate stage and the second candidate stage is less than the preset dispersion threshold (with a value of 0.05~0.1), the cloud platform will determine all three stages as joint root cause stages, forming a three-stage joint root cause determination result.

[0143] When the root cause stage is a single stage, the cloud platform selects the workstation with the highest causal contribution from all workstations included in that stage, and identifies it as the root cause workstation for the target PCB board defect. When the root cause stage is a joint root cause stage, the cloud platform selects the workstation with the highest causal contribution from each joint root cause stage, and identifies it as the joint root cause workstation. Based on the root cause workstation and the deviation weights of the feature types of each workstation within that workstation, the cloud platform further determines the root cause process parameters: the process parameters corresponding to the feature type of the workstation with the highest deviation weight within the root cause workstation are determined as the root cause process parameters for the target PCB board defect.

[0144] This embodiment achieves automated determination from the causal contribution of a workstation to the root cause stage through a three-level cascaded determination rule. The dominant stage determination rule is applicable to typical defect scenarios dominated by a single stage and can quickly locate the root cause.

[0145] In addition, this application provides a defect tracing system for PCBA surface mount lines, including: PCBA surface mount lines and a cloud platform connected to each workstation in the PCBA surface mount lines;

[0146] The PCBA (Printed Circuit Board Assembly) line is used to process target PCBs. When a defect type is detected in the target PCB during the processing, it sends the defect type to the cloud platform. The cloud platform, upon receiving the defect type from the PCBA line, uses a pre-set unique identifier to obtain cross-station feature data of the target PCB during its flow through the PCBA line. Based on the defect type and a pre-built historical cross-station feature database, it obtains a historical sample set. Using the cross-station feature data, historical sample set, and a pre-set causal contribution statistical algorithm, it obtains the causal contribution of each station in the PCBA line to the target PCB defect. Finally, based on the causal contribution of each station in the PCBA line to the target PCB defect, it traces the defect back to its source at each station.

[0147] Furthermore, the PCBA surface mount line includes: solder paste printing station, solder paste printing inspection station, chip mounting station, chip mounting position inspection station, reflow oven, and soldering quality inspection station arranged sequentially according to the process flow. The cloud platform is also used to generate process data for each station in the PCBA surface mount line when it receives process parameters input by the user, and then distribute this data to the corresponding station. The station process data includes solder paste printing process data, printing inspection process data, chip mounting station process data, chip mounting position inspection process data, reflow oven temperature control process data, and soldering quality inspection process data. The solder paste printing station is used to perform solder paste printing on the target PCB board based on the received solder paste printing process data. The solder paste printing inspection station is used to inspect the target PCB board based on the received printing inspection process data. The system includes: a solder paste printing station to check the solder paste printing quality and obtain target solder paste printing inspection data; a placement station to perform placement processing on the target PCB board based on received placement station process data; a placement position detection station to check the placement position of the target PCB board based on received placement position detection process data and obtain placement position detection data; a reflow oven to perform soldering processing on the target PCB board based on received reflow oven temperature control process data; and a soldering quality inspection station to check the soldering quality of the target PCB board based on received soldering quality inspection process data and obtain soldering quality inspection data. When the soldering of the target PCB board is determined to be unqualified, the station obtains the defect type of the target PCB board based on the soldering quality inspection data and uploads the defect type to the cloud platform.

[0148] In addition, the PCBA surface mount line also includes: a laser marking machine set between the automatic board suction station and the solder paste printing station; the laser marking machine is used to mark a unique identifier at a preset position on the target PCB board.

[0149] Furthermore, the PCBA surface mount line also includes a central control display device to show the status of the PCBA surface mount line during operation, and its tracking interface is as follows: Figure 2 As shown.

[0150] In the description of this application, although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for defect tracing of a PCBA surface mount line, the method being realized based on a pre-deployed PCBA surface mount line and a cloud platform, the PCBA surface mount line comprising a plurality of workstations arranged in a process flow direction in sequence, characterized in that, The method includes: When the cloud platform receives the defect type of the target PCB board from the PCBA surface mounting line, the cloud platform obtains the cross-station feature data of the target PCB board in the PCBA surface mounting line flow process based on the unique identifier pre-set in the target PCB board. The cloud platform obtains a historical sample set based on the defect type of the target PCB board and a pre-built historical cross-workstation feature database; wherein, the historical sample set includes historical cross-workstation feature data corresponding to multiple normal PCB boards and historical cross-workstation feature data corresponding to multiple defective PCB boards with the same defect type as the target PCB board. The cloud platform uses cross-station feature data of the target PCB board and the historical sample set, as well as a pre-set causal contribution statistical algorithm, to obtain the causal contribution of each station in the PCBA surface mounting line to the defects of the target PCB board. The causal contribution statistical algorithm is used to determine the causal contribution of each station in the PCBA surface mounting line to the defects of the target PCB board based on the distribution of cross-station feature data of the target PCB board in the historical sample set. The cloud platform traces the defects of the target PCB board based on the causal contribution of each workstation in the PCBA surface mount line to the defects of the target PCB board.

2. The defect tracing method for PCBA surface mount lines according to claim 2, characterized in that, The PCBA surface mount line includes: a solder paste printing station, a solder paste printing inspection station, a chip mounting station, a chip mounting position inspection station, a reflow oven, and a soldering quality inspection station arranged sequentially according to the process flow; When the cloud platform receives the defect type feedback from the PCBA surface mount line of the target PCB board, the cloud platform obtains the cross-station characteristic data of the target PCB board during the PCBA surface mount line flow process based on the pre-set unique identifier of the target PCB board, including: When the cloud platform receives the defect type of the target PCB board from the welding quality inspection station, the cloud platform, based on the unique identifier pre-set for the target PCB board, obtains the corresponding solder paste printing process data, solder paste printing inspection data, SMT station process data, SMT position detection data, reflow oven temperature control process data, and welding quality inspection data for the target PCB board. The solder paste printing process data refers to the process data of the solder paste printing station during solder paste printing on the target PCB board; the solder paste printing inspection data refers to the inspection data of the solder paste printing inspection station; the SMT station process data refers to the process data of the SMT station during SMT assembly on the target PCB board; the SMT position detection data refers to the SMT position data detected by the SMT position detection station; the reflow oven temperature control process data refers to the process data of the reflow oven during welding on the target PCB board; and the welding quality inspection data refers to the welding quality data detected by the welding quality inspection station. The cloud platform extracts features from solder paste printing process data, solder paste printing inspection data, chip mounting station process data, chip mounting position inspection data, reflow oven temperature control process data, and soldering quality inspection data, and sorts the quantified features of all extracted data according to the process flow to obtain cross-station feature data of the target PCB board.

3. The defect tracing method for surface-mount lines in PCBA according to claim 1, characterized in that, The cross-workstation feature data of the target PCB board and each of the aforementioned historical cross-workstation feature data includes multiple workstation feature types; The cloud platform obtains a historical sample set based on the defect types of the target PCB board and a pre-built historical cross-workstation feature database, including: Based on the defect type of the target PCB board, the cloud platform filters out the historical cross-station feature data corresponding to all defective PCB boards with the same defect type as the target PCB board from the pre-built historical cross-station feature database, as well as the historical cross-station feature data corresponding to all normal PCB boards without defects; and marks all defective PCB boards with defect labels. The cloud platform uses a pre-set equal-frequency binning discretization algorithm to discretize each selected historical cross-workstation feature data. The equal-frequency binning discretization strategy includes: for each workstation feature type, statistically analyzing the feature values ​​corresponding to that workstation feature type in all historical cross-workstation feature data to obtain the feature value distribution corresponding to each workstation feature type; and based on the feature value distribution corresponding to each workstation feature type, dividing each workstation feature type into a preset number of discrete intervals so that the feature values ​​corresponding to each workstation feature type in all historical cross-workstation feature data are mapped to the discrete intervals. A historical sample set is constructed based on all historical cross-station feature data after discrete processing and all defect labels.

4. The defect tracing method for PCBA surface mount lines according to claim 3, characterized in that, The cloud platform, based on cross-workstation feature data of the target PCB board and the historical sample set, as well as a pre-set causal contribution statistical algorithm, obtains the causal contribution of each workstation in the PCBA surface mounting line to the defects of the target PCB board, including: The cloud platform maps the feature values ​​of each workstation feature type in the cross-workstation feature data of the target PCB board to the corresponding discrete interval. The cloud platform, based on all historical cross-station feature data and all defect type labels in the historical sample set, obtains the joint probability distribution, marginal distribution of station features, and marginal distribution of defect labels for each feature type in the historical sample set. Specifically, the joint probability distribution represents the probability distribution of feature values ​​within each discrete interval for historical cross-station feature data corresponding to a particular station feature type, where the feature values ​​are either present or absent from the data. The marginal distribution of station features represents the probability distribution of feature values ​​within each discrete interval for historical cross-station feature data corresponding to a particular station feature type. The marginal distribution of defect labels represents the probability distribution of data with or without defect labels within all historical cross-station feature data corresponding to a particular station feature type. The cloud platform obtains mutual information corresponding to each workstation feature type based on the joint probability distribution, normal joint probability distribution, workstation feature marginal distribution, and defect label marginal distribution corresponding to each workstation feature type in the historical sample set. The cloud platform obtains the causal contribution of each workstation in the PCBA surface mounting line to the defects of the target PCB board based on the cross-workstation feature data of the target PCB board and the mutual information corresponding to the feature type of each workstation.

5. The defect tracing method for PCBA surface mount lines according to claim 4, characterized in that, The cloud platform obtains mutual information corresponding to each workstation feature type based on the joint probability distribution, normal joint probability distribution, workstation feature marginal distribution, and defect label marginal distribution corresponding to each workstation feature type in the historical sample set. This information includes: The cloud platform obtains the mutual information between each workstation feature type and the defect type based on the joint probability distribution, marginal distribution of workstation features, and marginal distribution of defect labels corresponding to each workstation feature type in the historical sample set, as well as a pre-set formula (Formula 1). Formula 1 is: ； where r(X i is the mutual information between the feature type of the workstation indexed by i and the defect type, k is the index of the discrete interval, K is the number of discrete intervals, and d is the index of whether the label of the defect is present. p(k) is the probability that the feature value falls into the kth discrete interval in the marginal distribution of the feature type of the workstation, p(d) is the probability that the label of the defect is present / absent in the marginal distribution of the label of the defect corresponding to the feature type of the workstation, and p(k,d) is the probability that the label of the defect is present / absent and the feature value is in the kth discrete interval in the joint probability distribution corresponding to the feature type of the workstation.

6. The defect tracing method for PCBA surface mount lines according to claim 4, characterized in that, The cloud platform, based on the cross-workstation feature data of the target PCB board and the mutual information corresponding to the feature type of each workstation, obtains the causal contribution of each workstation in the PCBA surface mounting line to the defects of the target PCB board, including: The cloud platform normalizes the mutual information corresponding to all workstation feature types to obtain the basic causal contribution degree corresponding to each workstation feature type; wherein, the sum of the basic causal contribution degrees corresponding to all workstation feature types is 1. The cloud platform obtains the deviation weight corresponding to each workstation feature type based on the cross-workstation feature data corresponding to the target PCB board and the historical cross-workstation feature data corresponding to all normal PCB boards in the historical sample set, as well as a pre-set formula two; the formula two is: ； Among them, w i x is the deviation weight corresponding to the feature type of the workstation with index i. i,目标板 Let i be the feature value of the cross-station feature data corresponding to the target PCB board at the station feature type with index i. This represents the mean of the feature values ​​for all normal PCB boards at the workstation feature type with index i. The standard deviation of the eigenvalues ​​for all normal boards on the feature type of the workstation at index i; The cloud platform, based on the deviation weight corresponding to the feature type of each workstation and the pre-set workstation feature table, performs a weighted summation of all basic causal contributions of each workstation in the PCBA surface mounting line to obtain the causal contribution of each workstation in the PCBA surface mounting line to the defect of the target PCB board; wherein, the workstation feature table is a correspondence table between workstations and workstation feature types in the PCBA surface mounting line.

7. The defect tracing method for surface mount lines in PCBA according to claim 2, characterized in that, The cloud platform performs workstation-based source tracing of defects on the target PCB board based on the causal contribution of each workstation in the PCBA surface mount line to the defects of the target PCB board, including: The cloud platform obtains upstream causal contribution, midstream causal contribution, and downstream causal contribution based on the causal contribution of each station in the PCBA surface mount line to the defects of the target PCB board. Among them, the upstream causal contribution includes the causal contribution of the solder paste printing station and the solder paste printing inspection station to the defects of the target PCB board; the midstream causal contribution includes the causal contribution of the chip mounting station and the chip mounting position inspection station to the defects of the target PCB board; and the downstream causal contribution includes the causal contribution of the reflow oven and the soldering quality inspection station to the defects of the target PCB board. The cloud platform determines the root cause stage of the target PCB board based on upstream causal contribution, midstream causal contribution, and downstream causal contribution, as well as a pre-set root cause stage judgment strategy, so as to perform workstation tracing of defects in the target PCB board.

8. A defect tracing system for surface mount lines in PCBA, characterized in that, include: PCBA surface mount cable and cloud platform connected to each workstation in the PCBA surface mount cable. The PCBA surface mount line is used to process the target PCB board and, when a defect type of the target PCB board is detected during the processing, sends the defect type of the target PCB board to the cloud platform. The cloud platform is used to obtain cross-station feature data of the target PCB board during the PCBA surface mounting process based on the unique identifier pre-set in the target PCB board when it receives the defect type feedback of the target PCB board. Based on the defect type of the target PCB board and a pre-built historical cross-station feature database, a historical sample set is obtained; wherein, the historical sample set includes historical cross-station feature data corresponding to multiple normal PCB boards and historical cross-station feature data corresponding to multiple defective PCB boards with the same defect type as the target PCB board. Based on the cross-station feature data of the target PCB board and the historical sample set, as well as the pre-set causal contribution statistical algorithm, the causal contribution of each station in the PCBA surface mounting line to the defects of the target PCB board is obtained; the causal contribution statistical algorithm is used to determine the causal contribution of each station in the PCBA surface mounting line to the defects of the target PCB board according to the distribution of the cross-station feature data of the target PCB board in the historical sample set. Based on the causal contribution of each workstation in the PCBA surface mount line to the defects of the target PCB board, the defects of the target PCB board are traced back to their source at each workstation.

9. The defect tracing system for PCBA surface mount lines according to claim 8, characterized in that, The PCBA surface mount line includes: a solder paste printing station, a solder paste printing inspection station, a chip mounting station, a chip mounting position inspection station, a reflow oven, and a soldering quality inspection station arranged sequentially according to the process flow; The cloud platform is also used to generate process data for each workstation in each PCBA surface mount line when it receives process parameters input by the user, and to send the data to the corresponding workstation. The process data includes solder paste printing process data, printing inspection process data, chip mounting workstation process data, chip mounting position detection process data, reflow oven temperature control process data, and soldering quality inspection process data. Solder paste printing station is used to perform solder paste printing on the target PCB board according to the received solder paste printing process data; The solder paste printing inspection station is used to inspect whether the solder paste printing on the target PCB board is qualified based on the received printing inspection process data, and to obtain the target solder paste printing inspection data. The placement station is used to perform placement processing on the target PCB board according to the received placement station process data; The chip placement detection station is used to detect whether the chip placement position of the target PCB board is qualified based on the received chip placement position detection process data, and to obtain chip placement position detection data. A reflow oven is used to perform soldering on a target PCB board based on received reflow oven temperature control process data. The welding quality inspection station is used to inspect whether the welding of the target PCB board is qualified based on the received welding quality inspection process data, and to obtain welding quality inspection data. When it is determined that the welding of the target PCB board is unqualified, the station obtains the defect type of the target PCB board based on the welding quality inspection data, and uploads the defect type to the cloud platform.

10. The defect tracing system for PCBA surface mount lines according to claim 9, characterized in that, The PCBA surface mount line also includes: a laser marking machine located between the automatic board suction station and the solder paste printing station; The laser marking machine is used to mark a unique identifier at a preset position on the target PCB board; The solder paste printing station is also used to obtain the unique identifier of the target PCB board and to map the solder paste printing process data to the unique identifier of the target PCB board. The solder paste printing inspection station is also used to obtain the unique identifier of the target PCB board and to map the target solder paste printing inspection data to the unique identifier of the target PCB board. The placement station is also used to obtain the unique identifier of the target PCB board and to map the placement station's process data to the unique identifier of the target PCB board. The chip placement location detection station is also used to obtain the unique identifier of the target PCB board and to map the chip placement location detection process data to the unique identifier of the target PCB board. The reflow oven is also used to obtain the unique identifier of the target PCB board and to map the reflow oven temperature control process data to the unique identifier of the target PCB board. The welding quality inspection station is also used to obtain the unique identifier of the target PCB board and to map the welding quality inspection data to the unique identifier of the target PCB board. The cloud platform is also used to obtain the mapping processing results of each workstation in the PCBA surface mounting line in real time.

Images