An industrial pipeline multi-source data processing method, system and medium based on asynchronous communication

Abstract

This invention provides a method, system, and medium for multi-source data processing in industrial production lines based on asynchronous communication. The method includes: receiving multi-source data from events at different workstations on the industrial production line via an asynchronous event queue; sorting events according to their priority and timestamps; performing multi-threaded parallel processing of the data using a backpressure mechanism and a persistence mechanism; performing coarse alignment of the data; performing time-series matching using a dynamic event warping algorithm; and introducing Kalman filtering and adaptive interpolation algorithms to correct lost or delayed data; performing anomaly detection and data cleaning using statistical methods, machine learning methods, and spatial consistency constraints; correcting the scan numbers of different workstations and establishing a mapping relationship between the corrected numbers and the data; performing data compression and storage; and performing parameter data visualization, trend warning, and anomaly prediction. This invention achieves accurate collection, integration, and traceability of data from different workstations, using different protocols, and with different timestamps.

Description

Technical Field

[0001] This invention relates to the field of industrial automation and data acquisition and processing technology, and in particular to a method, system and medium for multi-source data processing in industrial production lines based on asynchronous communication. Background Technology

[0002] In modern industrial production lines, common inspection and labeling processes include multi-angle CCD camera imaging, electrical performance testing, laser engraving, barcode scanning, and airtightness testing. These workstations often employ different communication protocols; for example, some use Modbus TCP, while others rely on RS232 / RS485 I / O communication.

[0003] Existing technologies suffer from a series of problems, including asynchronicity, differences in communication protocols, insufficient data quality, difficulty in tracing, and inadequate storage and utilization, as detailed below: 1. Asynchronousness issue: The different operating rhythms of each station on the production line make it impossible to strictly align the data collection time, which can easily lead to data misalignment or missing data. 2. Differences in communication protocols: With multiple workstations and multiple protocols coexisting, existing technologies struggle to achieve unified data acquisition and processing. 3. Insufficient data quality: Issues such as noise, outliers, and packet loss are common, leading to errors in the result tables; 4. Difficulty in data traceability: The lack of a reliable mapping between collected data and barcode numbers makes it difficult to ensure data correspondence throughout the entire product process; 5. Insufficient storage and utilization: Traditional methods often store data in simple table format, lacking compression, mining, and visualization capabilities.

[0004] Therefore, a highly reliable solution is needed that is compatible at the protocol layer and can achieve data alignment, cleaning, and visualization at the algorithm layer. Summary of the Invention

[0005] The main objective of this invention is to provide an asynchronous-driven, multi-protocol compatible, and algorithm-enhanced method, system, and medium for multi-source data processing in industrial production lines based on asynchronous communication. The aim is to achieve accurate collection, integration, and traceability of data from different workstations, different protocols, and different timestamps through an asynchronous triggering mechanism, a unified numbering mapping strategy, a multi-protocol adaptation layer, and an intelligent data processing module.

[0006] To achieve the above objectives, this invention proposes a multi-source data processing method for industrial production lines based on asynchronous communication, the method comprising the following steps: Step S10: Receive multi-source data from different workstations on the industrial production line through an asynchronous event queue, sort the events according to their priority and timestamps, and perform multi-threaded parallel processing of the data by combining backpressure and persistence mechanisms. Step S20: A sliding time window is used to coarsely align the data, and a dynamic event warping algorithm is used for time-series matching. Kalman filtering and adaptive interpolation algorithms are introduced to correct lost or delayed data in order to improve alignment accuracy. Step S30: Perform multi-dimensional anomaly detection and data cleaning by combining statistical methods, machine learning methods, and spatial consistency constraints; Step S40: Correct the scanning numbers of different workstations using a number correction method that combines a hash function and a minimum edit distance algorithm, and establish a mapping relationship between the corrected numbers and the data; Step S50: Data compression and storage are performed using Δ encoding and Huffman compression techniques, and parameter data visualization, trend warning, and anomaly prediction are performed using the LSTM / GRU time series prediction algorithm.

[0007] A further technical solution of the present invention is that, in step S10, the step of sorting events according to their priority and the timestamp of occurrence includes: The priority scheduling algorithm dynamically adjusts the priority of each event based on its base priority and waiting time. The formula for calculating the effective priority is: in, It is the final priority of the event, which determines the order in which the events are processed; It is the basic priority of events, determined by the event type; It is the waiting time impact coefficient, which indicates the degree to which waiting time affects event priority; It is the time that an event waits in the queue; Step S10 further includes: For each asynchronous event from different workstations on the industrial assembly line, a monotonic event stamp and a wall clock time are recorded simultaneously. The monotonic timestamp is used to ensure that the events are arranged in chronological order, and the wall clock time represents the absolute time when the event actually occurred. The steps for multi-threaded parallel processing of data by combining backpressure and persistence mechanisms include: The number of threads in the thread pool is dynamically adjusted according to the load, and events are distributed to threads in the multi-thread pool for parallel processing. When the load exceeds a preset threshold, non-critical events are delayed and stored. Step S10 further includes: Cache recent The event indexes in milliseconds form an event cache queue. Each event in the event cache queue has a status field and a confidence level. The status field is used to mark the current processing status of the event, and the confidence level is used to mark the reliability of the event.

[0008] A further technical solution of the present invention is that the step of using a sliding time window to coarsely align the data in step S20 includes: Step S201: Set the window width, for each event within the window width. Retrieve candidate events with similar timestamps from the event cache queue. and the event As related to the event Aligned candidate data; The step of performing time-series matching in step S20, which combines the dynamic event warping algorithm, includes: Step S202: Calculate the cumulative distance between the two sequences using Euclidean distance or absolute difference as the metric. Establish a cumulative distance matrix based on the cumulative distance of each point in the two sequences. Backtrack the cumulative matrix based on the optimal cost of dynamic event regularization to obtain the optimal matching path between the two sequences. Step S202 further includes: Step S203: By adding bandwidth constraints to limit the matching range, the search space is reduced and the calculation speed of the dynamic event warping algorithm is accelerated; In step S20, the steps of introducing Kalman filtering and adaptive interpolation algorithms to correct lost or delayed data in order to improve alignment accuracy include: Step S204: Kalman filtering is used to perform noise filtering, short missing data compensation, and delay correction on continuous quantities using the state-space method to achieve high-precision data alignment. The noise covariance is adjusted by online estimation to improve filtering accuracy and response speed. Step S205: In the case of missing data, the adaptive interpolation algorithm is used to accurately complete the long-term missing data by combining Kalman filter prediction and spline interpolation methods, ensuring the smoothness and accuracy of the data. During the interpolation process, the interpolation parameters are adjusted according to the residual estimation and data changes to optimize the interpolation accuracy.

[0009] A further technical solution of the present invention is that step S30 includes: Step S301: Outliers in the event dataset are identified using the Z-score algorithm, a statistical method. Data points that are greater than or equal to a preset threshold are identified as outliers; or outliers are identified by calculating the interquartile range (IQR) of the event data using the IOR algorithm. Step S302: For high-dimensional data, multiple trees are generated using a tree-based anomaly detection algorithm based on machine learning to isolate outliers and output anomaly scores, where the anomaly score represents the confidence level of the outlier. For normal samples, a One-Class SVM model based on machine learning is used to place normal samples on one side of the hyperplane and outliers on the other side. The output decision function value is mapped to a probability. If the probability is greater than a preset threshold, it is determined to be an anomaly. Step S303: Based on spatial consistency constraints, compare features from different CCD images to remove artifacts and erroneously acquired data.

[0010] A further technical solution of the present invention is that step S30 further includes: ensuring the timely removal of abnormal data based on adaptive threshold adjustment, abnormal recording and alarm mechanisms.

[0011] A further technical solution of the present invention is that step S40 includes performing fast hashing and bucketing on the scanned numbers, then using the minimum edit distance algorithm to calculate the normalized similarity, and correcting or manually verifying it by using a set threshold.

[0012] A further technical solution of the present invention is that step S40 further includes batch switching and parallel mapping: when a new batch enters the production line, the number sequence is automatically switched, and the data and number of the new batch are continued to be correctly matched; data acquisition and number mapping are performed on multiple workstations at the same time to ensure that the data and product number at each workstation can correspond to each other.

[0013] A further technical solution of the present invention is that step S40 further includes conflict detection and number verification: when different workstations collect the same number, it is marked as a conflict and an alarm mechanism is triggered to resolve the conflict; through historical record verification, each number is traced to ensure consistency and uniqueness in the production process.

[0014] To achieve the above objectives, the present invention also proposes an industrial production line multi-source data processing system based on asynchronous communication. The system includes a memory, a processor, and an industrial production line multi-source data processing program based on asynchronous communication stored on the processor. The industrial production line multi-source data processing program based on asynchronous communication is executed by the processor to perform the steps of the method described above.

[0015] To achieve the above objectives, the present invention also proposes a computer-readable storage medium storing an asynchronous communication-based multi-source data processing program for industrial production lines, wherein the asynchronous communication-based multi-source data processing program is executed by the processor to perform the steps of the method described above.

[0016] The beneficial effects of the present invention, which is a multi-source data processing method, system, and medium for industrial production lines based on asynchronous communication, are as follows: 1. Strong compatibility: Supports multiple protocols such as Modbus TCP, RS232, and RS485, adaptable to both new and old machines; 2. High robustness of asynchronous operation: The asynchronous event queue and global numbering mechanism ensure accurate data correspondence between different workstations. 3. High level of intelligence: The scanning and detection data are automatically bound, and the algorithm alignment and anomaly removal improve data reliability; 4. Excellent real-time performance: Millisecond-level data acquisition, number mapping, and data alignment meet the cycle time requirements of high-speed pipelines; 5. Strong traceability: Number mapping + compressed storage + visual reports ensure that the data of each product is traceable throughout the entire process; 6. High security and fault tolerance: Anomaly detection, packet loss retry, and alarm mechanisms ensure stable system operation; 7. Scalability and predictive capability: The LSTM / GRU trend prediction and adaptive algorithm can be extended to multi-station and multi-pipeline applications. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a preferred embodiment of the multi-source data processing method for industrial production lines based on asynchronous communication according to the present invention. Figure 2 This is a schematic diagram of the overall process of the multi-source data processing method for industrial production lines based on asynchronous communication according to the present invention.

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0020] This invention proposes a multi-source data processing method for industrial production lines based on asynchronous communication. The technical solution mainly uses an asynchronous triggering mechanism, a unified numbering mapping strategy, a multi-protocol adaptation layer, and intelligent data processing to achieve accurate collection, integration, and traceability of data from different workstations, different protocols, and different event stamps.

[0021] like Figure 1 and Figure 2 As shown, a preferred embodiment of the multi-source data processing method for industrial production lines based on asynchronous communication of the present invention includes the following steps: Step S10: Receive multi-source data from different workstations on the industrial production line through an asynchronous event queue, sort the events according to their priority and timestamps, and perform multi-threaded parallel processing of the data by combining backpressure and persistence mechanisms.

[0022] This invention relates to an asynchronous communication-based multi-source data processing method for industrial production lines. This method is applied to an asynchronous communication-based multi-source data processing system for industrial production lines. In this system, in order to efficiently process asynchronous events from different workstations and protocols, this invention adopts an asynchronous communication and acquisition mechanism, combining asynchronous event queues, priority scheduling algorithms, and multi-threaded parallel processing technologies to ensure priority processing of critical events (such as defects and alarms) and to ensure the stability of the system under high load.

[0023] An asynchronous event queue is used to store and manage asynchronous events from different workstations and using different protocols. Whenever an event occurs, it is added to the asynchronous event queue, awaiting further processing. To efficiently handle these events, the system needs to sort them according to their priority and timestamp, and process them according to a scheduling strategy.

[0024] Among them, the event queue ( ) represents the queue of all pending asynchronous events, timestamp ( This is used to record the time when an event occurs, which is then used to determine the event's position in the queue.

[0025] The priority scheduling algorithm determines the processing order of events by calculating the effective priority of each event. The system dynamically adjusts the priority of events based on their base priority and waiting time to ensure that important events (such as defects and alarms) are processed first.

[0026] Specifically, in this embodiment, step S10, which involves sorting events according to their priority and the timestamp of their occurrence, includes: The priority scheduling algorithm dynamically adjusts the priority of each event based on its base priority and waiting time. The formula for calculating the effective priority is: in, It is the final priority of the event, which determines the order in which the events are processed; It is the basic priority of events, which is usually determined by the event type. For example, defects have a higher priority than alarms. This is the waiting time impact coefficient, which indicates the degree to which waiting time affects event priority. A larger coefficient indicates a lower impact. The value indicates that the waiting time has a significant impact on the priority. This is the time an event waits in the queue; the longer the wait time, the higher the priority may be.

[0027] This scheduling strategy ensures that important events are prioritized even under high load conditions.

[0028] To improve event processing efficiency, this embodiment employs multithreaded parallel processing technology. Each event is assigned to a thread or coroutine and processed in parallel by multiple thread pools, ensuring that the system can run efficiently and stably even under high load.

[0029] Thread pool: This indicates the number of threads available in the thread pool. The system allocates and manages tasks by pre-creating multiple thread pools.

[0030] Task duration: It is the time required to process each event, which depends on the complexity of the event and the allocation of system resources.

[0031] During multi-threaded parallel processing, the system adjusts the thread pool size based on the load to achieve optimal resource utilization efficiency. Furthermore, the system also enables a backpressure mechanism based on the load; when the system load is too high, some non-critical events may be delayed to ensure that important events are processed with priority.

[0032] This embodiment employs a backpressure mechanism to ensure that no important events are lost when the system load reaches its limit. By restricting the entry of new events or temporarily postponing the processing of certain events, the backpressure mechanism ensures the stability of the system.

[0033] This embodiment employs a persistence mechanism, where events can be stored in a local persistence buffer or database until the system has sufficient resources to process them.

[0034] In this embodiment, step S10 further includes: For each asynchronous event from different workstations on the industrial assembly line, a monotonic event stamp and a wall clock time are recorded simultaneously. The monotonic timestamp is used to ensure that the events are arranged in chronological order, and the wall clock time represents the absolute time when the event actually occurred.

[0035] In this embodiment, step S10, which combines the backpressure mechanism and the persistence mechanism to perform multi-threaded parallel processing of data, includes: The number of threads in the thread pool is dynamically adjusted according to the load, and events are distributed to threads in the multi-thread pool for parallel processing. When the load exceeds a preset threshold, non-critical events are delayed and stored.

[0036] In this embodiment, step S10 further includes: Cache recent The event indexes in milliseconds form an event cache queue. Each event in the event cache queue has a status field and a confidence level. The status field is used to mark the current processing status of the event, and the confidence level is used to mark the reliability of the event.

[0037] In this embodiment, the process of step S10 is as follows: Events are added to the queue: When an event occurs, it is added to the asynchronous event queue.

[0038] Priority sorting: Calculate the effective priority based on the base priority and waiting time of each event, and sort the events according to the effective priority.

[0039] Multi-threaded parallel processing: Events are assigned to threads in a multi-threaded pool for parallel processing. The number of threads in the pool is dynamically adjusted according to the load to ensure that high-priority events are processed first.

[0040] Backpressure and persistence: When the load is too high, the backpressure mechanism is activated, and some non-critical events are delayed in processing. At the same time, the events are persisted until the system is able to process them.

[0041] Event handling and feedback: After the event is handled, the result will be transmitted to the system through a feedback mechanism to ensure a timely response.

[0042] By combining asynchronous event queues, priority scheduling algorithms, and multi-threaded parallel processing, this embodiment can efficiently handle asynchronous events from multiple workstations and protocols, ensuring system stability and response speed even under extreme loads.

[0043] This embodiment introduces a timestamp caching and data marking mechanism to address issues of duplicate and missed data acquisition in high-speed pipelines, thereby improving data real-time performance and robustness. Each asynchronous event records two timestamps simultaneously in the system: a monotonic timestamp (…). ) and wall clock time ( Monotonic timestamps ensure events are arranged in chronological order, while wall clock times represent the absolute time when the event actually occurred. Multi-station concurrent acquisition and asynchronous buffering are supported to ensure overall data integrity across the pipeline.

[0044] To improve system processing efficiency and reliability, the system caches recently used data in memory. The event indexes in milliseconds form an event cache queue. Each event also includes a status field (...). ) and confidence level ( This is used to mark the current processing status of events (e.g., "pending," "aligned") and the reliability of the events (the degree of data alignment). During event alignment, the system determines whether events match by calculating the time difference between them. Specifically, if events... and events Time difference satisfy: in, It is an event and events The time difference between two events. It represents the absolute difference in time between two events.

[0045] and These are events and events The wall-clock time is the actual time at which they occur, representing the system's true time.

[0046] It represents the time difference between two events, and calculates the absolute difference in their occurrence times. That is, regardless of which event occurs first, the calculation result is always positive, ensuring that the alignment judgment is not affected by the order.

[0047] This is the alignment tolerance time window, representing the maximum allowed time difference. If the time difference between two events is less than or equal to this value, the system considers them alignable. This value is typically set to a few milliseconds, depending on the real-time requirements of the system.

[0048] Step S20: A sliding time window is used to coarsely align the data, and a dynamic event warping algorithm is used for time-series matching. Kalman filtering and adaptive interpolation algorithms are introduced to correct lost or delayed data in order to improve alignment accuracy.

[0049] For asynchronous data such as multi-angle shooting from CCD cameras, electrical measurement, and airtightness detection, this embodiment uses the sliding time window and dynamic time warping (DTW) method to efficiently align the time sequence of data from different sensors, ensuring data consistency and accuracy.

[0050] In this embodiment, the step of using a sliding time window to coarsely align the data in step S20 includes: Step S201: Set the window width, for each event within the window width. Retrieve candidate events with similar timestamps from the event cache queue. and the event As related to the event Aligned candidate data.

[0051] The core idea of ​​the sliding time window mechanism is to set a window width ( Within this window, events are coarsely aligned. The purpose of this window is to determine the search range for candidate data, thereby performing initial alignment of data with small time differences and reducing the computational load of the subsequent Dynamic Time Warping (DTW) algorithm. Specifically, for each event... (Hypothetical event) With timestamp The system will retrieve candidate events from the cache that have a similar timestamp. And use it as candidate data for alignment. Indicates an event timestamp; This indicates the window width and the length of the time window. Generally, a setting of 50ms is recommended, but the specific value can be adjusted according to the device's clock speed.

[0052] In this embodiment, the step of performing time-series matching in step S20 using the dynamic event warping algorithm includes: Step S202: Calculate the cumulative distance between the two sequences using Euclidean distance or absolute difference as the metric. Establish a cumulative distance matrix based on the cumulative distance of each point in the two sequences. Backtrack the cumulative matrix based on the optimal cost of dynamic event regularization to obtain the optimal matching path between the two sequences.

[0053] Dynamic Time Warping (DTW) is a method for time alignment that calculates the cumulative distance between two sequences. DTW aims to find the time alignment between two sequences (…). and The optimal matching path between () is determined by considering their potential nonlinear time deformation.

[0054] Equipment sequence and : :sequence ,Include Data points.

[0055] :sequence ,Include Data points.

[0056] Local distance : Represents a sequence Points in and sequence Points in The distance between them. Typically, Euclidean distance or absolute difference is chosen as the metric. in, It is a distance metric function, and its value is usually Euclidean distance or absolute difference.

[0057] Cumulative distance matrix Dynamic time warping constructs a cumulative distance matrix. To store the cumulative distance of each pair of matching points, the size of the matrix is... Each element Indicates from sequence and From the starting point to and The cumulative distance.

[0058] DTW optimal cost The final optimal cost of DTW is a matrix. The bottom right element in the text, i.e. , indicating from sequence to sequence The minimum twisted path cost.

[0059] Matching path By backtracking the cumulative distance matrix, the optimal matching path between two sequences can be obtained. Each matching point... Represents a sequence The Middle Data points and sequences The Middle The best match for each data point.

[0060] In this embodiment, step S202 further includes: Step S203 involves adding bandwidth constraints to limit the matching range, thereby reducing the search space and accelerating the computation speed of the dynamic event warping algorithm.

[0061] To improve computational efficiency, Dynamic Time Warping (DTW) algorithms typically incorporate bandwidth constraints to limit the range of matches. This embodiment employs the Sakoe-Chiba bandwidth constraint, which reduces the search space by limiting the index difference between matching points, thereby accelerating computation.

[0062] The bandwidth constraint formula is: in, This is a bandwidth parameter, representing the maximum allowed index difference. Through bandwidth constraints, the system only considers... and The difference does not exceed The point pairs significantly reduce computational complexity.

[0063] To address the high computational complexity of DTW, this embodiment utilizes the FastDTW algorithm, which accelerates DTW execution through approximate calculations. Furthermore, bandwidth... The settings can further limit the calculation range and reduce unnecessary calculations.

[0064] The normalized cost formula is: After calculating the normalization cost, the system compares it with a preset threshold to determine the matching relationship between events.

[0065] Furthermore, in this embodiment, step S20, which introduces Kalman filtering and adaptive interpolation algorithms to correct lost or delayed data in order to improve alignment accuracy, includes: Step S204: Kalman filtering is used to perform noise filtering, short missing data compensation, and delay correction on continuous quantities using the state-space method to achieve high-precision data alignment. The noise covariance is adjusted by online estimation to improve filtering accuracy and response speed. Step S205: In the case of missing data, the adaptive interpolation algorithm is used to accurately complete the long-term missing data by combining Kalman filter prediction and spline interpolation methods, ensuring the smoothness and accuracy of the data. During the interpolation process, the interpolation parameters are adjusted according to the residual estimation and data changes to optimize the interpolation accuracy.

[0066] This embodiment introduces Kalman filtering combined with an adaptive interpolation algorithm to correct missing or delayed data, achieving millisecond-level alignment accuracy. A state-space method is used to perform noise filtering, short-missing data compensation, and delay correction on continuous quantities to improve alignment accuracy (target: millisecond-level comparability). Kalman filtering achieves high-precision data alignment by using a state-space method to perform noise filtering, short-missing data compensation, and delay correction on continuous quantities, and adjusts the noise covariance through online estimation. and To improve filtering accuracy and response speed, an adaptive interpolation algorithm combines Kalman filtering prediction and spline interpolation methods to accurately complete long-term missing data in the case of missing data, ensuring data smoothness and accuracy.

[0067] 1) Kalman Filtering The Kalman filter is a linear filter based on a state-space model, widely used for noise filtering, missing data compensation, and delay correction for continuous signals. Its goal is to optimally estimate state variables based on system predictions and observations, correcting for missing or delayed data in the process, thereby achieving millisecond-level alignment accuracy.

[0068] The Kalman filter formula and parameter meanings are as follows: In standard form, the state and observation equations of the Kalman filter are as follows: Equations of state: in, It is a state vector that contains the current state variables of the system, such as signal values ​​and their first-order differences (velocity, etc.). It is the state transition matrix, used to describe the dynamic changes of the system. It is the control input, which represents the influence of external inputs on the system state. It is process noise, usually assumed to be Gaussian noise, with a covariance of... .

[0069] The observation equation is as follows: in, It is an observation vector, representing the system's observed values ​​or measurement results. It is the observation matrix, used to map state variables to the observation space. It is observation noise, usually assumed to be Gaussian noise, with a covariance of... .

[0070] Process noise covariance matrix and observation noise covariance matrix : It is the covariance matrix of the process noise, reflecting the uncertainty of the system's dynamic model. It is the covariance matrix of the observation noise, representing the uncertainty of the observation data.

[0071] The Kalman filtering steps are as follows: A. Prediction steps: Predict the current state vector and estimate the current state based on the state and control input from the previous time step.

[0072] B. Update steps: in, The Kalman gain determines the weight distribution between prediction and observation. It is the predicted covariance matrix, which reflects the uncertainty of the state.

[0073] Adaptive adjustment and : Adjusted by online residual estimation (Observation noise covariance) The noise covariance is updated using the residual between the actual observations and the predicted values.

[0074] Adjustment through state prediction error This ensures a balance between filtering accuracy and response speed.

[0075] 2) Adaptive Interpolation Algorithm Adaptive interpolation algorithms are used to correct missing data, especially in the case of long-term missing data in time series. By using prediction and interpolation techniques, lost data information can be recovered, further improving the accuracy of data alignment.

[0076] The adaptive interpolation algorithm steps are as follows: For missing data points, the system first predicts the state at the missing location using Kalman filtering, and then uses the predicted values ​​to imput the missing data points. For short-term missing data, the system performs simple imputation based on the predicted state.

[0077] For data missing for extended periods, the system employs a smoothing algorithm or spline interpolation (such as B-spline interpolation) to complete the data. Spline interpolation fits the data using a polynomial function, ensuring the smoothness and continuity of the interpolation results.

[0078] Spline interpolation: Spline interpolation is used to complete data that has been missing for a long time. It uses higher-order polynomials or piecewise polynomials for interpolation to ensure the smoothness of the data.

[0079] Suppose we have a set of data points The goal is to construct a smooth interpolation function using these points. .

[0080] B-spline interpolation approximates the data by constructing a low-order polynomial (such as a quadratic or cubic polynomial) over each interval, ensuring that the interpolation curve is continuous and has a consistent derivative at each node.

[0081] Adaptive update: For each interpolation, the algorithm automatically adjusts the interpolation parameters based on the residual estimate and changes in the data to optimize the interpolation accuracy.

[0082] For CCD images, feature matching algorithms (SURF / ORB) combined with timestamp association can be used to ensure that data captured from multiple angles corresponds to the same product; visual features are used to enhance the alignment confidence across workstations / sensors. Feature matching and timestamp association: Image feature matching is performed using ORB descriptors, Hamming distance, and the Lowe ratio test, combined with timestamp association to ensure accurate alignment of data across workstations and sensors. Multidimensional association matrix and matching algorithm: A multidimensional association matrix containing time, features, and ID is constructed, and optimal data matching is performed through normalized weights and matching algorithms to ensure unified management and efficient integration of data across workstations.

[0083] 1) Feature matching algorithm and timestamp association In the image data processing, ORB (Oriented Fast and Rotated BRIEF) descriptors and Hamming distance are used for feature matching, while timestamp association is combined to enhance alignment confidence across workstations and sensors. This method ensures that CCD image data taken from different angles can be accurately mapped to the same product, thereby improving the accuracy of data alignment.

[0084] ORB descriptors and matching: ORB descriptors: An ORB descriptor is a binary feature vector used to describe key points in an image. Each descriptor... For one A binary vector of bits, usually .

[0085] This indicates that each descriptor consists of 256 bits of binary data.

[0086] Hamming distance: In ORB matching, the similarity between two descriptors is determined by calculating their Hamming distance. The Hamming distance is the bit difference between two binary vectors. in and These are descriptors and In the The value of the Hamming distance. The smaller the Hamming distance, the more similar the two descriptors are, and the higher the probability of a match.

[0087] Lowe ratio test: This test is used to filter out false matches. For each descriptor Find its nearest neighbor and next nearest neighbor If the distance ratio between the nearest neighbor and the second nearest neighbor is less than a certain threshold (e.g., 0.8), then a match is accepted. This method effectively reduces false matches.

[0088] Timestamp association: To ensure data alignment across different sensors or workstations, the system uses timestamp association to match image data and sensor data. timestamp difference : Calculate the time difference between the time the image is captured and the time the sensor data is acquired.

[0089] Represents the timestamp of the image.

[0090] The timestamp represents the sensor data.

[0091] Image matching candidate conditions: if time difference Less than or equal to the set time window If the image matches, then it is considered a candidate. If the time difference exceeds the window If the match is not found, then discard the match.

[0092] Result: By associating ORB descriptors with timestamps, we can ensure that data from different angles or different sensors can be aligned and effectively matched.

[0093] 2) Multidimensional association matrix and matching algorithm During image alignment and data matching, the system generates a multidimensional correlation matrix, comprehensively considering time, feature, and ID information to achieve unified data management across workstations, providing reliable input for subsequent analysis. By fusing time, feature, and ID information into a unified cost or confidence matrix, the system can employ matching algorithms (such as the Hungarian algorithm, maximum weight matching, or greedy algorithm) to achieve optimal data matching.

[0094] Confidence (score) model: Model formula: in: It is a normalized weight of time, feature, and ID, satisfying . It is a score based on time difference (e.g., using an exponential decay model). It is a normalized confidence level based on dynamic time warping (DTW) or image matching. It is the score for matching IDs, with 1 for the same ID. Levenshtein similarity is calculated based on the similarity of IDs.

[0095] Time-difference-based scoring (exponential decay): in and These are timestamps for two events. It is a smoothing coefficient that controls the impact of time difference on the matching score.

[0096] Hungarian Algorithm and Matching: Cost Matrix: Construct a multi-dimensional cost matrix that includes time, features, and ID. Each element Indicates an event and events The matching cost or confidence level between them.

[0097] Matching algorithm: For small-scale problems, the Hungarian algorithm can be used for globally optimal matching, with a time complexity of O(n log n). .

[0098] For high-throughput scenarios, temporal partitioning and greedy approximation are used, and threshold pruning is used to reduce the number of candidate matches to ensure real-time performance.

[0099] Virtual node completion: If the set and If the sizes of the nodes are inconsistent, virtual nodes are used to pad them to form a square matrix. The matching cost of the virtual nodes is set to a large constant, allowing for "unmatched" nodes.

[0100] Final match: The system calculates matching scores using normalized weights, prioritizes the matching pairs with the highest scores, and establishes cross-workstation mappings based on the matching results. For high-throughput real-time scenarios, the system uses an approximation algorithm to ensure a fast response in the matching process.

[0101] Step S30 involves combining statistical methods (Z-score, IQR), machine learning methods (Isolation Forest, One-Class SVM), and spatial consistency constraints to perform multi-dimensional anomaly detection and data cleaning, detecting and removing anomalous data to ensure high-precision data cleaning.

[0102] In this embodiment, step S30 includes...

[0103] Step S301: Outliers in the event dataset are identified using the Z-score algorithm, a statistical method. Data points that are greater than or equal to a preset threshold are identified as outliers; or outliers are identified by calculating the interquartile range (IQR) of the event data using the IOR algorithm.

[0104] The Z-score algorithm, a statistical method, is used to identify outliers in a dataset. A data point is considered an anomaly when its absolute Z-score value exceeds a certain threshold.

[0105] The formula used in the Z-score algorithm of statistical methods is as follows: in, It's a data point. It is the sample mean. That is the sample standard deviation. If... If the threshold is not met, the data point is considered an anomaly. A common threshold is... .

[0106] The IQR method identifies outliers by calculating the interquartile range (IQR) of data. Outliers are typically located in... or In addition, among them and These are the first quartile and the third quartile, respectively.

[0107] Step S302: For high-dimensional data, multiple trees are generated using a tree-based anomaly detection algorithm based on machine learning methods to isolate outliers and output anomaly scores, where the anomaly score represents the confidence level of the outlier. For normal samples, a One-Class SVM model based on machine learning methods is used to place normal samples on one side of the hyperplane and outliers on the other side. The output decision function value is mapped to a probability. If the probability is greater than a preset threshold, it is determined to be an anomaly.

[0108] The Isolation Forest algorithm in machine learning is a tree-based anomaly detection method suitable for high-dimensional data. This algorithm "isolates" outliers by generating multiple trees, ultimately outputting an anomaly score. The score is Within the range, it represents the confidence level of outliers.

[0109] Its decision-making criteria are based on the set contamination (proportion of abnormal samples) or threshold to determine whether the data is abnormal.

[0110] The One-Class SVM algorithm in machine learning is used to identify cases where only normal samples exist. By training the model, most normal samples are positioned on one side of a hyperplane, while abnormal samples fall on the other. Its output decision function value... It can be mapped to probability using the sigmoid function: If the probability is greater than a certain set threshold, it is judged as abnormal.

[0111] Step S303: Based on spatial consistency constraints, compare features from different CCD images to remove artifacts and erroneously acquired data.

[0112] In this embodiment, spatial consistency constraints are implemented by comparing features from different CCD images to eliminate artifacts and erroneously acquired data. By comparing multiple images, data consistency is ensured and errors are reduced.

[0113] In this embodiment, step S30 further includes: ensuring the timely removal of abnormal data based on adaptive threshold adjustment, abnormal recording, and alarm mechanisms.

[0114] Based on the above-mentioned anomaly detection, this embodiment adopts an adaptive threshold adjustment and alarm mechanism to ensure the timely removal of abnormal data and the stable operation of the system.

[0115] In this embodiment, the adaptive threshold adjustment is mainly based on automatic optimization of filter parameters. Specifically, the filter parameters are automatically adjusted according to changes in the pipeline environment to adapt to different working environments. Through online learning and feedback mechanisms, this embodiment can dynamically adjust the threshold and parameters to optimize anomaly detection results.

[0116] This embodiment records all detected anomalies and provides queryable logs to ensure subsequent traceability and auditing.

[0117] For high-confidence anomalies, the system will automatically trigger an alarm to remind operators or relevant personnel to handle the anomaly in a timely manner. Alarms include, but are not limited to, data loss, sensor drift, false data, etc.

[0118] In this embodiment, the anomaly scoring method is as follows: assuming the anomaly scores for each method are respectively... (Statistical methods) (Machine learning methods) and (Spatial consistency), final anomaly score For weighted average: in, The weights of each method satisfy the following conditions: .

[0119] In this embodiment, the anomaly detection scheme is as follows: Threshold determination: If Greater than the set threshold If the data is abnormal, an alarm will be triggered or the process will begin manual review.

[0120] Step S40: Correct the scanning numbers of different workstations using a number correction method that combines a hash function and a minimum edit distance algorithm, and establish a mapping relationship between the corrected numbers and the data.

[0121] In this embodiment, step S40 includes performing fast hashing and bucketing on the scanned numbers, then using the minimum edit distance algorithm to calculate the normalized similarity, and correcting or manually verifying it by using a set threshold.

[0122] Furthermore, in this embodiment, step S40 also includes batch switching and parallel mapping: when a new batch enters the production line, the number sequence is automatically switched, and the data and number of the new batch are continued to be correctly matched; data collection and number mapping are performed on multiple workstations simultaneously to ensure that the data and product number at each workstation can correspond consistently.

[0123] Furthermore, in this embodiment, step S40 also includes conflict detection and number verification: when different workstations collect the same number, it is marked as a conflict and an alarm mechanism is triggered to resolve the conflict; through historical record verification, each number is traced to ensure consistency and uniqueness in the production process.

[0124] To ensure the correctness and consistency of the serial numbers under asynchronous acquisition, this embodiment uses a combination of hash functions and the Levenshtein algorithm to automatically correct the scanned serial numbers. The process includes standardizing the scanned serial numbers, fast hashing and binning, then using the Levenshtein algorithm to calculate the normalized similarity, and performing automatic correction or manual verification based on a set threshold.

[0125] 1) QR code number correction and hashing algorithm with minimum edit distance Standardized QR code numbering: Standardization: the process of standardizing the strings obtained from the scan. The code performs standardization, removing spaces, unifying capitalization, and mapping characters. For example, it replaces the letter "O" with the number "0".

[0126] Standardized operation: This process ensures that the same number can maintain a consistent format under different scanning conditions.

[0127] Hash functions and bucketing: Fast hashing: This method uses a hash function (such as CRC32 or FNV) to hash the normalized ID, generating a hash value and assigning the ID to different buckets, thus reducing the number of matching candidates. The goal of the hash function is to group similar IDs into the same bucket.

[0128] The hash formula is: The numbers are assigned to buckets based on their hash values, and only numbers within the same bucket will be subject to subsequent minimum edit distance calculations.

[0129] Minimum edit distance (Levenshtein distance): Levenshtein's algorithm: used to calculate the minimum edit distance between two strings, that is, the minimum number of operations (including insertion, deletion, and replacement) required to transform one string into another.

[0130] Recursive definition: Represents a string and The minimum edit distance. The cost of character replacement, when It is 1 if it is true, otherwise it is 0.

[0131] Normalized similarity: The minimum edit distance calculated using Levenshtein is used to calculate the normalized similarity. in, For strings and The minimum edit distance between them. This similarity value is between 0 and 1, with higher similarity values ​​closer to 1.

[0132] Automatic correction and manual review: Automatic correction: If normalized similarity Greater than the set threshold (like If the two numbers match, then the system will automatically correct the error. Manual review: If the similarity is below the threshold, the match will be manually reviewed to ensure the correctness of the number.

[0133] 2) Numbering Mapping and Conflict Detection During production, the system not only needs to ensure the accuracy of the data number for each product, but also needs to handle batch switching, multi-station parallel mapping, and conflict detection. By generating historical data on the numbering and verifying the mapping, the system can provide accurate production traceability and quality audit support.

[0134] Batch switching and parallel mapping: Batch switching: During production, the system supports batch switching. When a new batch enters the production line, the system automatically switches the numbering sequence and continues to ensure that the data and numbers of the new batch match correctly.

[0135] Multi-station parallel mapping: The system supports multiple workstations to collect data and map numbers simultaneously, ensuring that the data at each workstation corresponds to the product number.

[0136] Conflict detection and number verification: Conflict Detection: Through real-time monitoring and verification, the system can detect potential numbering conflicts. For example, when two workstations collect the same number, the system will mark it as a conflict and trigger an alarm mechanism. Conflict resolution methods include modifying the numbering and reassigning the workstation.

[0137] Number verification: Through historical verification, the system can trace each number to ensure its consistency and uniqueness in the production process.

[0138] Number history and mapping log: Numbering History: The system records the change history of all numbers for quality auditing and traceability analysis. Each number change generates a log, recording the time and reason for the change.

[0139] Mapping Log: A log is generated for each mapping process with each number, which facilitates subsequent quality control and manual review.

[0140] Step S50: Data compression and storage are performed using Δ encoding and Huffman compression techniques, and parameter data visualization, trend warning, and anomaly prediction are performed using the LSTM / GRU time series prediction algorithm.

[0141] 1) Δ encoding and Huffman compression During data acquisition, Δ coding and Huffman compression techniques are combined to reduce storage pressure and improve storage efficiency. Δ coding reduces redundant data by storing the differences in the time series, while Huffman coding further compresses the data by constructing an optimal prefix code.

[0142] Δ encoding (differential encoding): Δ coding is a method of encoding time series data. Its basic idea is to store the difference between each data point and the previous data point. This method is particularly effective for data with relatively small variations because the differencing data is usually more concentrated and has less variation than the original data, making it suitable for entropy coding and compression.

[0143] The Δ encoding formula is as follows: Given a time series its difference sequence for: in It is the first The difference between each data point and its preceding data point. By storing the difference sequence instead of the original sequence, the storage requirement can be effectively reduced, especially when the data changes little.

[0144] Huffman Coding: Huffman coding is a lossless data compression algorithm that compresses data by constructing an optimal prefix code. Its core idea is to assign shorter codes to frequently occurring symbols and longer codes to less frequently occurring symbols.

[0145] The steps of Huffman coding are as follows: Constructing a frequency table: Given a set of symbols and the probability of each symbol First, calculate the frequency of each symbol.

[0146] Calculate information entropy: Information entropy The formula used to measure the average information content of a set of symbols is: in It is a symbol The probability, This represents the average information content of the set.

[0147] Constructing the Huffman tree: A minimum priority queue is built based on the probability of the symbols, and the Huffman tree is constructed by merging pairs of symbols with the lowest frequency. The root node of the tree represents the entire dataset, and the leaf nodes represent individual symbols.

[0148] Prefix code generation: Optimal prefix codes are generated using Huffman trees, with shorter codes assigned to more frequent symbols and longer codes assigned to less frequent symbols.

[0149] Compression process: Each symbol is replaced with its corresponding binary prefix code, thereby achieving data compression.

[0150] 2) Time series forecasting and visualization After data acquisition and processing, the system can generate trend curves, distribution histograms, and defect rate prediction curves in real time. Furthermore, the system can also enable... LSTM (Long Short-Term Memory) network or GRU (Gated Recurrent Unit) Time series forecasting algorithms are used to provide trend warnings and anomaly predictions for key parameters.

[0151] LSTM (Long Short-Term Memory Network): LSTM is a special type of recurrent neural network (RNN) used to process and predict time series data. LSTM has a unique gating mechanism that allows it to maintain long-term dependencies during training.

[0152] The basic unit of LSTM consists of an input gate, a forget gate, and an output gate. Its input and state update formulas are as follows: in, This is the current input. It is the hidden state of the previous time step. These are the activation values ​​for the forget gate, input gate, and output gate, respectively. It is a cellular state.

[0153] The state update formula is as follows: GRU (Gated Cyclic Unit): GRU is a simplified version of LSTM, which combines the forget gate and input gate in LSTM, reducing the amount of computation.

[0154] GRU basic unit: It's the update gate, which determines what percentage of the current state comes from historical states. It's a reset gate, controlling the forgetting of past information.

[0155] Forecasting and Early Warning: Time series forecasting: LSTM / GRU can be used to perform short-term forecasting and trend extrapolation of time series data, helping the system to perform trend warning and anomaly detection.

[0156] Prediction results: Through the model's output, the system can identify abnormal trends or fluctuations in key parameters in advance, thereby providing early warnings.

[0157] The beneficial effects of the multi-source data processing method for industrial production lines based on asynchronous communication in this invention are: 1. Strong compatibility: Supports multiple protocols such as Modbus TCP, RS232, and RS485, adaptable to both new and old machines; 2. High robustness of asynchronous operation: The asynchronous event queue and global numbering mechanism ensure accurate data correspondence between different workstations. 3. High level of intelligence: The scanning and detection data are automatically bound, and the algorithm alignment and anomaly removal improve data reliability; 4. Excellent real-time performance: Millisecond-level data acquisition, number mapping, and data alignment meet the cycle time requirements of high-speed pipelines; 5. Strong traceability: Number mapping + compressed storage + visual reports ensure that the data of each product is traceable throughout the entire process; 6. High security and fault tolerance: Anomaly detection, packet loss retry, and alarm mechanisms ensure stable system operation; 7. Scalability and predictive capability: The LSTM / GRU trend prediction and adaptive algorithm can be extended to multi-station and multi-pipeline applications.

[0158] To achieve the above objectives, the present invention also proposes an industrial production line multi-source data processing system based on asynchronous communication. The system includes a memory, a processor, and an industrial production line multi-source data processing program based on asynchronous communication stored on the processor. The industrial production line multi-source data processing program based on asynchronous communication is executed by the processor to perform the steps of the method described above.

[0159] To achieve the above objectives, the present invention also proposes a computer-readable storage medium storing an asynchronous communication-based multi-source data processing program for industrial production lines, wherein the asynchronous communication-based multi-source data processing program is executed by the processor to perform the steps of the method described above.

[0160] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural changes made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. An asynchronous communication based industrial pipeline multi-source data processing method, characterized in that, The method includes the following steps: Step S10: Receive multi-source data from different workstations on the industrial production line through an asynchronous event queue, sort the events according to their priority and timestamps, and perform multi-threaded parallel processing of the data by combining backpressure and persistence mechanisms. Step S20: A sliding time window is used to coarsely align the data, and a dynamic event warping algorithm is used for time-series matching. Kalman filtering and adaptive interpolation algorithms are introduced to correct lost or delayed data in order to improve alignment accuracy. Step S30: Perform multi-dimensional anomaly detection and data cleaning by combining statistical methods, machine learning methods, and spatial consistency constraints; Step S40: Correct the scanning numbers of different workstations using a number correction method that combines a hash function and a minimum edit distance algorithm, and establish a mapping relationship between the corrected numbers and the data; Step S50: Data compression and storage are performed using Δ encoding and Huffman compression techniques, and parameter data visualization, trend warning, and anomaly prediction are performed using the LSTM / GRU time series prediction algorithm.

2. The multi-source data processing method for industrial production lines based on asynchronous communication according to claim 1, characterized in that, In step S10, the step of sorting events according to their priority and the timestamp of occurrence includes: The priority scheduling algorithm dynamically adjusts the priority of each event based on its base priority and waiting time. The formula for calculating the effective priority is: in, It is the final priority of the event, which determines the order in which the events are processed; It is the basic priority of events, determined by the event type; It is the waiting time impact coefficient, which indicates the degree to which waiting time affects event priority; It is the time the event waits in the queue; Step S10 further includes: For each asynchronous event from different workstations on the industrial assembly line, a monotonic event stamp and a wall clock time are recorded simultaneously. The monotonic timestamp is used to ensure that the events are arranged in chronological order, and the wall clock time represents the absolute time when the event actually occurred. The steps for multi-threaded parallel processing of data by combining backpressure and persistence mechanisms include: The number of threads in the thread pool is dynamically adjusted according to the load, and events are distributed to threads in the multi-thread pool for parallel processing. When the load exceeds a preset threshold, non-critical events are delayed and stored. Step S10 further includes: Cache recent The event indexes in milliseconds form an event cache queue. Each event in the event cache queue has a status field and a confidence level. The status field is used to mark the current processing status of the event, and the confidence level is used to mark the reliability of the event.

3. The multi-source data processing method for industrial production lines based on asynchronous communication according to claim 1, characterized in that, The step of coarsely aligning the data using a sliding time window in step S20 includes: Step S201: Set the window width, for each event within the window width. Retrieve candidate events with similar timestamps from the event cache queue. and the event As related to the event Aligned candidate data; The step of performing time-series matching in step S20, which combines the dynamic event warping algorithm, includes: Step S202: Calculate the cumulative distance between the two sequences using Euclidean distance or absolute difference as the metric. Establish a cumulative distance matrix based on the cumulative distance of each point in the two sequences. Backtrack the cumulative matrix based on the optimal cost of dynamic event regularization to obtain the optimal matching path between the two sequences. Step S202 further includes: Step S203: By adding bandwidth constraints to limit the matching range, the search space is reduced and the calculation speed of the dynamic event warping algorithm is accelerated; In step S20, the steps of introducing Kalman filtering and adaptive interpolation algorithms to correct lost or delayed data in order to improve alignment accuracy include: Step S204: Kalman filtering is used to perform noise filtering, short missing data compensation, and delay correction on continuous quantities using the state-space method to achieve high-precision data alignment. The noise covariance is adjusted by online estimation to improve filtering accuracy and response speed. Step S205: In the case of missing data, the adaptive interpolation algorithm is used to accurately complete the long-term missing data by combining Kalman filter prediction and spline interpolation methods, ensuring the smoothness and accuracy of the data. During the interpolation process, the interpolation parameters are adjusted according to the residual estimation and data changes to optimize the interpolation accuracy.

4. The multi-source data processing method for industrial production lines based on asynchronous communication according to claim 1, characterized in that, Step S30 includes: Step S301: Outliers in the event dataset are identified using the Z-score algorithm, a statistical method. Data points that are greater than or equal to a preset threshold are identified as outliers; or outliers are identified by calculating the interquartile range (IQR) of the event data using the IOR algorithm. Step S302: For high-dimensional data, multiple trees are generated using a tree-based anomaly detection algorithm based on machine learning to isolate outliers and output anomaly scores, where the anomaly score represents the confidence level of the outlier. For normal samples, a One-Class SVM model based on machine learning is used to place normal samples on one side of the hyperplane and outliers on the other side. The output decision function value is mapped to a probability. If the probability is greater than a preset threshold, it is determined to be an anomaly. Step S303: Based on spatial consistency constraints, compare features from different CCD images to remove artifacts and erroneously acquired data.

5. The multi-source data processing method for industrial production lines based on asynchronous communication according to claim 4, characterized in that, Step S30 further includes: ensuring the timely removal of abnormal data based on adaptive threshold adjustment, abnormal recording and alarm mechanisms.

6. The multi-source data processing method for industrial production lines based on asynchronous communication according to claim 1, characterized in that, Step S40 includes performing fast hashing and bucketing on the scanned numbers, then using the minimum edit distance algorithm to calculate the normalized similarity, and correcting or manually verifying it by using a set threshold.

7. The multi-source data processing method for industrial production lines based on asynchronous communication according to claim 1, characterized in that, Step S40 also includes batch switching and parallel mapping: when a new batch enters the production line, the numbering sequence is automatically switched, and the data and numbering of the new batch are continued to be correctly matched; data collection and numbering mapping are performed on multiple workstations at the same time to ensure that the data and product number at each workstation can correspond to each other.

8. The multi-source data processing method for industrial production lines based on asynchronous communication according to claim 1, characterized in that, Step S40 further includes conflict detection and number verification: when different workstations collect the same number, it is marked as a conflict and an alarm mechanism is triggered to resolve the conflict; through historical record verification, each number is traced to ensure consistency and uniqueness in the production process.

9. A multi-source data processing system for industrial production lines based on asynchronous communication, characterized in that, The system includes a memory, a processor, and an asynchronous communication-based multi-source data processing program for industrial production lines stored on the processor, wherein the asynchronous communication-based multi-source data processing program is executed by the processor to perform the steps of the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an asynchronous communication-based industrial production line multi-source data processing program, which, when executed by the processor, performs the steps of the method as described in any one of claims 1 to 8.

Images