A production line defective product counter data flow processing method and system

Abstract

The application provides a production line defective product counter data flow processing method and system, and relates to the technical field of production line quality detection.The method comprises the following steps: collecting discrete defect trigger signals triggered by each quality inspection station of a production line, adding a multi-dimensional context identifier comprising a time stamp, a spatial coordinate and a type code to the discrete defect trigger signals to obtain an initial event vector; constructing a time sequence sliding window according to the initial event vector + time stamp, and calculating the data correlation strength between each event vector in the time sequence sliding window; constructing a dynamic coupling field model according to the data correlation strength, simulating the data fluctuation of the discrete defect trigger signals as a mechanical motion track, performing anti-shake deduplication and frequency domain limiting processing on the initial event vector by limiting the acceleration change amplitude of the data change rate to obtain a smooth and continuous structured event sequence.The application realizes flow smoothing monitoring and dynamic hierarchical early warning of production line defective product data.

Description

Technical Field

[0001] This invention relates to the field of production line quality inspection technology, and in particular to a method and system for processing data streams of a defective product counter on a production line. Background Technology

[0002] In modern manufacturing assembly line production, real-time monitoring of defective products is a crucial aspect of quality control. Currently, many production lines rely on electronic counters based on physical buttons or simple data acquisition terminals to record defective product information. Each time an operator discovers a defective product, they press the corresponding button, the equipment records a count, and uploads the data to a host computer. However, in practical applications, this existing counting method is often ill-suited to complex production environment fluctuations. For example, in high-speed assembly processes, when operators encounter a series of similar defective products, such as consecutive scratches, they may trigger the same physical button multiple times within a very short time interval due to haste or hand tremors.

[0003] Most existing counting devices only have simple signal accumulation functions and lack the ability to deeply analyze the timing characteristics of signals. Faced with the high-frequency trigger signals mentioned above, existing devices may not be able to effectively distinguish whether they are genuine continuous defective product outputs or accidental trigger signals caused by operational jitter. This results in a large amount of noise being mixed into the recorded data. More importantly, since most existing systems only focus on the accumulation of discrete count values ​​and ignore the continuity and rate of change characteristics of the data in the time dimension, the system cannot capture the accelerating trend of a rapid increase in the defect rate in a short period of time. It may be difficult to identify subtle changes or early signs of deterioration in the quality status of the production line in a timely manner, thus failing to provide effective intervention and early warning before a large-scale outbreak of defective products. Summary of the Invention

[0004] This invention provides a method and system for processing data streams of defective product counters on production lines, enabling smooth streaming monitoring and dynamic hierarchical early warning of defective product data on production lines.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: A first aspect includes a method for processing data streams from a production line defective product counter, the method comprising: Step 1: Collect discrete defect triggering signals triggered by each quality inspection station on the production line, add multi-dimensional context identifiers including timestamps, spatial coordinates and type codes to the discrete defect triggering signals, and obtain the initial event vector; Step 2: Construct a time-series sliding window based on the timestamps of the initial event vectors, and calculate the data correlation strength between each event vector within the time-series sliding window; construct a dynamic coupled field model based on the data correlation strength, simulate the data fluctuations of the discrete defect trigger signal as a mechanical motion trajectory, and perform anti-shake deduplication and frequency domain limiting processing on the initial event vectors by limiting the acceleration change amplitude of the data change rate, to obtain a smooth and continuous structured event sequence; Step 3: Perform multi-channel aggregation calculation on the structured event sequence to track the flow trajectory of each defect type in the process chain, and calculate the instantaneous change rate and trend acceleration of the cumulative defect amount; then perform multi-dimensional feature weighted fusion of the instantaneous change rate, trend acceleration and the flow trajectory of the defect type to obtain the quality state distribution tensor. Step 4: Perform phase boundary deduction on the quality state distribution tensor to determine the warning phase region of the current quality state. Correct the preset static critical threshold according to the warning phase region to obtain the dynamic critical threshold. Map the dynamic critical threshold to form a real-time quality topology dashboard and hierarchical warning control flow to realize streaming monitoring and dynamic hierarchical response of defective product count data on the production line.

[0006] Secondly, a data stream processing system for a production line defect counter includes: The acquisition and identification module is used to acquire discrete defect triggering signals triggered by each quality inspection station on the production line, and add multi-dimensional context identifiers including timestamps, spatial coordinates and type codes to the discrete defect triggering signals to obtain an initial event vector. The dynamic coupling module is used to construct a time-series sliding window based on the initial event vectors and calculate the data correlation strength between each event vector within the time-series sliding window. Based on the data correlation strength, a dynamic coupling field model is constructed to simulate the data fluctuations of discrete defect trigger signals as mechanical motion trajectories. By limiting the acceleration change amplitude of the data change rate, anti-shake deduplication and frequency domain limiting processing are performed on the initial event vectors to obtain a smooth and continuous structured event sequence. The aggregation calculation module is used to perform multi-channel aggregation calculations on structured event sequences, track the flow trajectory of each defect type in the process chain, and calculate the instantaneous change rate and trend acceleration of the cumulative defect amount. The instantaneous change rate, trend acceleration and the flow trajectory of the defect type are then fused using multi-dimensional feature weighting to obtain the quality state distribution tensor. The dynamic early warning module is used to perform phase boundary deduction on the quality state distribution tensor, determine the early warning phase region of the current quality state, and correct the preset static critical threshold according to the early warning phase region to obtain the dynamic critical threshold. Based on the dynamic critical threshold, a real-time quality topology dashboard and hierarchical early warning control flow are formed to realize the streaming monitoring and dynamic hierarchical response of the defective product count data of the production line.

[0007] The above-described solution of the present invention has at least the following beneficial effects: By adding multi-dimensional context identifiers such as timestamps, spatial coordinates, and type codes to discrete defect trigger signals, an initial event vector is constructed to achieve standardized and structured acquisition of defect signals from each quality inspection station on the production line. A time-series sliding window and a dynamic coupled field model are used to perform anti-jitter deduplication and frequency domain amplitude limiting processing on discrete signals, transforming discrete fluctuating data into a smooth and continuous structured event sequence, thus improving the accuracy and stability of the original defect count data. Multi-channel aggregation calculations are used to track the defect process flow trajectory, and multi-dimensional features such as instantaneous change rate and trend acceleration are integrated to form a quality state distribution tensor, dynamically depicting the propagation law and development trend of defects in the production line and identifying early signs of quality anomalies. Based on the quality state distribution tensor, phase boundaries are deduced and critical thresholds are dynamically corrected to adapt to the fluctuations in production line conditions and process changes. Combined with a real-time quality topology dashboard and hierarchical early warning control flow, streaming real-time monitoring and hierarchical response of defect count data are achieved. Attached Figure Description

[0008] Figure 1 This is a schematic flowchart of a data stream processing method for a defective product counter on a production line, provided by an embodiment of the present invention.

[0009] Figure 2 This is a schematic diagram of a production line defective product counter data stream processing system provided by an embodiment of the present invention. Detailed Implementation

[0010] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0011] like Figure 1 As shown, an embodiment of the present invention proposes a data stream processing method for a defective product counter on a production line, the method comprising the following steps: Step 1: Collect discrete defect triggering signals triggered by each quality inspection station on the production line, add multi-dimensional context identifiers including timestamps, spatial coordinates and type codes to the discrete defect triggering signals, and obtain the initial event vector; Step 2: Construct a time-series sliding window based on the timestamps of the initial event vectors, and calculate the data correlation strength between each event vector within the time-series sliding window; construct a dynamic coupled field model based on the data correlation strength, simulate the data fluctuations of the discrete defect trigger signal as a mechanical motion trajectory, and perform anti-shake deduplication and frequency domain limiting processing on the initial event vectors by limiting the acceleration change amplitude of the data change rate, to obtain a smooth and continuous structured event sequence; Step 3: Perform multi-channel aggregation calculation on the structured event sequence to track the flow trajectory of each defect type in the process chain, and calculate the instantaneous change rate and trend acceleration of the cumulative defect amount; then perform multi-dimensional feature weighted fusion of the instantaneous change rate, trend acceleration and the flow trajectory of the defect type to obtain the quality state distribution tensor. Step 4: Perform phase boundary deduction on the quality state distribution tensor to determine the warning phase region of the current quality state. Correct the preset static critical threshold according to the warning phase region to obtain the dynamic critical threshold. Map the dynamic critical threshold to form a real-time quality topology dashboard and hierarchical warning control flow to realize streaming monitoring and dynamic hierarchical response of defective product count data on the production line.

[0012] In this embodiment of the invention, by adding multi-dimensional context identifiers such as timestamps, spatial coordinates, and type codes to discrete defect trigger signals, an initial event vector is constructed to achieve standardized and structured acquisition of defect signals from each quality inspection station on the production line. A time-series sliding window and a dynamic coupled field model are used to perform anti-shake deduplication and frequency domain amplitude limiting processing on discrete signals, transforming discrete fluctuating data into a smooth and continuous structured event sequence, thus improving the accuracy and stability of the original defect count data. Multi-channel aggregation calculations are used to track the defect process flow trajectory, and multi-dimensional features such as instantaneous change rate and trend acceleration are integrated to form a quality state distribution tensor, dynamically depicting the propagation law and development trend of defects in the production line and identifying early signs of quality anomalies. Based on the quality state distribution tensor, phase boundaries are deduced and critical thresholds are dynamically corrected to adapt to the fluctuations in production line conditions and process changes. Combined with a real-time quality topology dashboard and hierarchical early warning control flow, streaming real-time monitoring and hierarchical response of defect count data are achieved.

[0013] In a preferred embodiment of the present invention, step 1 above, which involves collecting discrete defect triggering signals from each quality inspection station on the production line, adding a multi-dimensional context identifier including a timestamp, spatial coordinates, and type code to the discrete defect triggering signals to obtain an initial event vector, may include: In this embodiment of the invention, step 110 involves deploying defect detection sensors at each quality inspection station. When a defective product passes through the detection area, a discrete defect trigger signal in the form of an electrical pulse is obtained and transmitted to the front-end data acquisition unit to obtain the original pulse sequence. Specifically, this includes: in the operating area of ​​each quality inspection station on the production line, defect detection sensors are fixedly deployed according to industrial production assembly standards. The selected defect detection sensors are physical button-type defect trigger sensors adapted to high-speed assembly production lines. The sensors are deployed according to the number of defective products detected at each station, with each sensor corresponding to a single defective product detection type. The sensor installation height matches the operator's hand gestures, and the installation position is flush with the station's operating panel without obstruction to ensure smooth triggering. The sensor's signal output end is connected to the digital signal input port of the front-end data acquisition unit using an industrial-grade shielded twisted-pair cable. The shielded twisted-pair cable can resist the operation of the production line motor and electromagnetic interference. Environmental factors are considered to ensure signal transmission is distortion-free and attenuated. The defect detection sensor outputs a standard digital electrical pulse signal. In normal standby mode, it outputs a low level (0V). When the operator completes defect detection and triggers the sensor button, the sensor outputs a high level (3.3V) momentarily for 20 milliseconds. After one trigger, it automatically returns to a low level. This electrical pulse signal is the discrete defect trigger signal. The front-end data acquisition unit continuously collects the electrical pulse signals transmitted by each sensor at a fixed sampling frequency of 1 MHz. It captures and temporarily stores each received electrical pulse signal in real time. According to the order of signal reception, all captured electrical pulse signals are arranged and integrated to form a continuous and complete original pulse sequence. The original pulse sequence completely preserves the amplitude, duration, and reception timing characteristics of each defect trigger electrical signal, without signal truncation, distortion, or loss.

[0014] Step 111 involves capturing the rising or falling edge of each pulse in the original pulse sequence and adding a timestamp to each pulse to obtain a pulse event stream with timing identifiers. Specifically, this includes: the front-end data acquisition unit initiates a hardware-level pulse edge capture mechanism. This mechanism uses an 8 MHz high-precision crystal oscillator inside the acquisition unit as the clock source and achieves microsecond-level signal recognition accuracy through clock frequency division. It can stably capture both rising and falling edges of the pulse. Rising edge capture detects the instant the electrical pulse signal transitions from 0V low level to 3.3V high level, while falling edge capture detects the instant the electrical pulse signal transitions from 3.3V high level to 0V low level. In actual production scenarios, the rising edge capture mode is used by default to ensure that signal locking is completed immediately upon triggering the action. After capturing the edge of a single pulse, the front-end data acquisition unit immediately calls the built-in real-time clock module to obtain the precise time data of the current moment. The timing accuracy of the real-time clock module reaches the millisecond level. The time data includes year, month, date, hour, minute, second, and millisecond. This time data is used as a timing identifier and a one-to-one binding relationship is established with the currently captured pulse signal to ensure that each pulse signal corresponds to a unique trigger time. All pulse signals that have completed edge capture and are bound with timing identifiers are arranged in the order of capture to form a pulse event stream with timing identifiers. Each independent event in the pulse event stream contains three basic types of information: pulse edge type, original electrical signal characteristics, and timing identifier, which can clearly distinguish the order of triggering actions of different defects.

[0015] Step 112: Based on the quality inspection station number from which each pulse originates in the pulse event stream, retrieve the three-dimensional spatial coordinate values ​​of the station in the physical layout of the production line from a preset spatial coordinate mapping table, and add the three-dimensional spatial coordinate values ​​as spatial identifiers to the corresponding pulse events to obtain a positioning event stream with timestamps and spatial coordinates. Specifically, this includes: pre-building and storing a spatial coordinate mapping table within the front-end data acquisition unit. This mapping table is a fixed data table pre-configured before production line deployment. The pre-built process involves assigning consecutive integer numbers to all quality inspection stations according to the production line process flow sequence, starting from 1 and incrementing sequentially, with each station corresponding to a unique station number; and obtaining the three-dimensional spatial coordinates of each quality inspection station in the physical layout of the production line through precise on-site surveying. The three-dimensional spatial coordinates use millimeters as the uniform unit, where the X-axis represents the horizontal coordinate along the length of the production line. The Y-axis represents the horizontal coordinate in the width direction of the production line, and the Z-axis represents the vertical coordinate in the height direction of the workstation installation. A unique mapping relationship is established between the workstation number and the corresponding three-dimensional spatial coordinates, and the data is entered into a spatial coordinate mapping table. There are no duplicates, missing data, or erroneous mapping data in the table. When processing the pulse event stream, the front-end data acquisition device first extracts the quality inspection workstation number corresponding to each pulse signal. Using the workstation number as the retrieval index, a matching retrieval is performed in the preset spatial coordinate mapping table to retrieve the three-dimensional spatial coordinate value that uniquely corresponds to the workstation number. The retrieved X-axis coordinate, Y-axis coordinate, and Z-axis coordinate are used as spatial identifiers and attached to the corresponding pulse event. All pulse events are bound to spatial identifiers, maintaining the original temporal order. This is integrated to form a positioning event stream with timestamps and spatial coordinates. The positioning event stream can determine the physical location of each defect triggering action in the production line.

[0016] Step 113: Based on the currently set inspection items of the quality inspection station corresponding to each pulse event, assign a preset defect type code as a type identifier to each event in the location event stream and add it to the event to obtain a three-dimensional identified event stream with timestamp, spatial coordinates, and type code. Specifically, this includes: pre-building and storing a defect type code configuration table inside the front-end data acquisition unit. This configuration table is a fixed data table that is pre-configured according to the defective product classification standard of the production line before production. The pre-construction process is as follows: first, sort out all defective product types on the production line, including appearance scratches, dimensional deviations, functional failures, assembly misalignments, etc. Each defective product type corresponds to a unique inspection item; assign a unique decimal integer type code to each defective product type, starting from 1001. The codes are sequentially incremented, with no duplicate or missing codes. A one-to-one correspondence is established between defective product types, inspection items, and type codes, and this is entered into a defect type code configuration table to ensure that each inspection item can be directly matched to a fixed type code. When processing the location event stream, the front-end data collector reads the currently preset inspection items of the quality inspection station corresponding to each pulse event. Using the inspection items as the matching basis, the corresponding preset type code is retrieved from the defect type code configuration table, and this type code is attached as a type identifier to the corresponding event in the location event stream. After the type identifier is added, each pulse event integrates three types of contextual information: timestamp, three-dimensional spatial coordinates, and defect type code. All events carrying complete information are arranged in their original chronological order to form a three-dimensional identifier event stream.

[0017] Step 114: Each event in the 3D identified event stream is encapsulated sequentially into a structured data tuple containing three fields: timestamp, spatial coordinates, and type code, according to the order of its timestamps. This forms the initial event vector. Specifically, the front-end data collector performs temporal sorting on all independent events in the 3D identified event stream. The sorting uses a stable sorting algorithm, with the timestamp value attached to the event as the sole criterion. Assume the 3D identified event stream contains... There are 1 independent events, and any two events are denoted as _ ... and , The corresponding timestamp value is , The corresponding timestamp value is The sorting determination formula is: if < ,but Arranged in ahead; if = The events are sorted in ascending order according to their original receiving order, ensuring that the event sequence strictly follows the actual time sequence of defect triggering. After sorting, each independent event is encapsulated in a standardized structure, forming a data tuple with five fixed fields: timestamp, X-axis spatial coordinate, Y-axis spatial coordinate, Z-axis spatial coordinate, and defect type code. Each field stores the event's time sequence, spatial location, and defect type information. The data types and dimensions of the fields are standardized, with no format conflicts or dimension confusion. This encapsulated structured data tuple is the initial event vector. All initial event vectors are combined sequentially in chronological order to form the initial event vector sequence.

[0018] The defect trigger signal is transformed into a standardized initial event vector carrying three core information: timing, physical location, and defect type. This distinguishes between real defect triggers and invalid signals generated by operational jitter or accidental touches, thereby enhancing the traceability and early intervention capabilities for production line quality issues.

[0019] In a preferred embodiment of the present invention, step 2 above involves constructing a time-series sliding window from the initial event vectors according to timestamps, and calculating the data correlation strength between each event vector within the time-series sliding window; constructing a dynamic coupled field model based on the data correlation strength, simulating the data fluctuations of the discrete defect trigger signal as a mechanical motion trajectory, and performing anti-jitter deduplication and frequency domain limiting processing on the initial event vectors by limiting the acceleration change amplitude of the data change rate, thereby obtaining a smooth and continuous structured event sequence, which may include: In this embodiment of the invention, step 220 involves extracting the timestamp attached to each event vector from the initial event vectors and sorting all event vectors according to the order of the timestamps to obtain an event vector sequence arranged in ascending order of time. Specifically, this includes: the front-end data collector performing a full traversal of all encapsulated initial event vectors, extracting the standardized timestamp field value from each initial event vector. The initial event vectors are structured data tuples with a fixed structure, containing five core fields: timestamp, X-axis spatial coordinates, Y-axis spatial coordinates, Z-axis spatial coordinates, and defect type code. The timestamp field has been numerically converted, uniformly using milliseconds as the unique unit, and is a continuously comparable decimal value, which can be directly used for time series size determination and sorting operations. After extracting the timestamps of all event vectors, the front-end data collector starts a stable insertion sort algorithm adapted to the embedded system, using the timestamp value as the sole sorting criterion, and performs a strict ascending order sort on all initial event vectors. Let the total number of initial event vectors participating in the sorting be... Mark all initial event vectors sequentially as , , ... The timestamp values ​​corresponding to the event vectors are marked sequentially as follows: , , ... The sorting process follows fixed decision rules for any two independent event vectors. and Their timestamps are respectively and The sorting logic is as follows: < ,Will Fixed arrangement in The position in front ensures that events triggered earlier are prioritized; if = Strictly preserve the original hardware receiving order of the two event vectors, without swapping their positions, to avoid timing errors; if > ,Will Fixed arrangement in The order of events is determined by placing them at the back of the list, ensuring that later-triggering events are sorted sequentially. The entire sorting process is complete without any event vector loss, duplicate sorting, or temporal misalignment. After sorting all event vectors, a sequence of event vectors is generated that is strictly arranged in ascending order of time. This sequence perfectly matches the actual time sequence of production line defect triggers.

[0020] Step 221: Based on the preset window length and sliding step parameters, multiple consecutive time-series sliding windows are sequentially divided from the starting position on the event vector sequence arranged in ascending order of time. Each time-series sliding window covers a fixed time span or a fixed number of event vectors, resulting in a windowed event vector grouping sequence. Specifically, this includes: the front-end data collector calling the preset core configuration parameters of the time-series sliding window in the production deployment stage. The parameters support two industrial scenario adaptation modes: fixed time span mode and fixed number of events mode. All parameters are configurable fixed values ​​with unified units and no conflicts. Fixed time span mode (adapted to high-speed continuous production scenarios), preset window length parameter. =1000ms, preset sliding step size parameter =200ms, each time-series sliding window covers a continuous time range of 1 second, and a new sliding window is generated every 200 milliseconds; fixed event quantity mode (adapted to intermittent production scenarios), preset window length parameter. =8, meaning each window contains 8 consecutive event vectors, with a preset sliding step size parameter. =2, meaning that each slide moves the event vector backward by 2.

[0021] The front-end data collector starts from the beginning of the time-incrementing event vector sequence and performs continuous sliding partitioning according to preset parameters. In the fixed time span mode, the first event vector in the sequence is used as the starting time, and all event vectors within the range of the starting time to the starting time + 1000ms are extracted to form the first time-series sliding window. Then, it moves forward in 200ms increments and repeats the extraction operation until all event vectors in the sequence are covered. In the fixed number of events mode, starting from the first event vector in the sequence, 8 event vectors are extracted continuously to form the first time-series sliding window. Then, it moves forward 2 event vectors and repeats the extraction operation until the number of remaining event vectors in the sequence is less than the window length. All the partitioned time-series sliding windows are combined in the order of their generation to form a windowed event vector grouping sequence. Each group corresponds to an independent time-series sliding window. The event vectors within the window retain their original time-series order, with no overlap or omissions.

[0022] For the 3D spatial coordinates of the event vectors within each time-series sliding window, an elliptical axis alignment transformation algorithm is performed. The complete algorithm process is as follows: Extract the 3D spatial coordinates of all m event vectors within the current time-series sliding window after standardization and storage, forming a 3D spatial coordinate set. The X, Y, and Z axis coordinates are uniformly measured in millimeters, where m represents the actual number of event vectors within the current window. The mean center of the spatial coordinates is calculated based on all spatial coordinates within the window to solve for the mean center of the three-dimensional distribution. , , The center is the reference origin aligned with the ellipse axis, and the calculation formula is: , , ,in , , Let m be the three-dimensional spatial coordinates of the k-th event vector within the window, and m be the total number of event vectors within the window. The three-dimensional covariance matrix is ​​constructed with the mean center as the coordinates. , , Based on this, a three-dimensional covariance matrix describing the correlation of spatial coordinate distributions is constructed. The matrix is ​​a 3×3 symmetric matrix, and each element in the matrix... The calculation formula is: ,in For the first Each event vector The coordinate values ​​of the axis. The mean center is at The coordinate values ​​of the axis. For the first One defective product incident The original spatial coordinate values ​​on the coordinate axes for All events within the current window Mean values ​​of coordinates on the coordinate axes (coordinates of the center of space). The correction coefficients are used for unbiased estimation; the eigenvalues ​​and orthogonal eigenvectors of the covariance matrix are obtained by performing eigenvalue decomposition on the three-dimensional covariance matrix Σ to obtain three non-negative eigenvalues. And three orthogonal unit eigenvectors corresponding one-to-one with the eigenvalues. The largest eigenvalue Corresponding feature vector The second largest eigenvalue is located along the major axis of the spatially distributed ellipse. Corresponding feature vector For the direction of the minor axis of the ellipse, the minimum eigenvalue is... Corresponding feature vector With the spatial normal direction as the reference direction, the three eigenvectors form an orthogonal reference coordinate system for axis-aligned transformation; the spatial coordinate axis-aligned projection transformation transforms the original three-dimensional spatial coordinates of each event vector within the window. Project the coordinates onto the orthogonal reference coordinate system formed by the principal axes of the ellipse, complete the axis alignment transformation, and the transformed 3D coordinates are: The conversion formula is: ,in This is the transpose of the eigenvectors, used for coordinate projection calculation transformation. After completion, all spatial coordinates are regularly distributed along the principal axis of the ellipse.

[0023] Step 222: For each event vector group within a time-series sliding window, extract any two event vectors from that group sequentially, and calculate the timestamp difference, Euclidean distance between their spatial coordinates, and type encoding matching coefficient to form a multidimensional difference vector. Specifically, for each event vector group within a time-series sliding window that has completed elliptical axis alignment transformation, the front-end data collector performs a full combination of double event vector traversal, sequentially extracting any two different event vectors from the group, denoted as the target event vector. and comparison event vectors The three core difference parameters—timestamp difference, spatial Euclidean distance, and type coding matching coefficient—are calculated separately. The specific calculation process is as follows: timestamp difference calculation, extraction... timestamp and timestamp Calculate the absolute time difference between two event vectors. The unit of measurement is millisecond, and the calculation formula is: , The smaller the value, the shorter the time interval between the two defect triggering events; the larger the value, the longer the time interval; spatial coordinate Euclidean distance calculation, extraction 3D coordinates after elliptical axis alignment transformation ,as well as Transformed 3D coordinates Calculate the three-dimensional Euclidean distance between two events in the physical space of the production line. The unit of measurement is uniformly millimeters, and the calculation formula is: , The smaller the value, the closer the physical locations of the two defect triggering events; the larger the value, the farther apart the physical locations. Defect type coding matching coefficient calculation, extraction... Defect type coding and Defect type coding Define type encoding matching coefficient The coefficient is a binary discrete value, and the determination rule is as follows: , =1 indicates that the two events are defects of the same type. =0 indicates that the two events are different types of defects; the multidimensional difference vector is constructed by dividing the timestamp difference. Spatial Euclidean distance Type encoding reverse difference value (1- This is integrated to form a three-dimensional multidimensional difference vector that represents the comprehensive differences between the two event vectors in the three dimensions of time sequence, space, and defect type. The vector expression is The multidimensional difference vector fully quantifies the degree of difference between two event vectors across all dimensions, with no missing information and no dimension conflict.

[0024] Step 223: Based on the multidimensional difference vector, calculate a value between 0 and 1 according to the preset weighted summation rules. This value is used as the data correlation strength between the two event vectors. Specifically, the front-end data collector calls the preset multidimensional feature normalization weight parameters before production. The weights are set according to the production line quality control priority. The time-series dimension is the core control dimension, the spatial dimension is the secondary control dimension, and the type dimension is the auxiliary control dimension. The three weights satisfy the normalization constraint condition, that is, the sum of the weights is always equal to 1. The specific preset value is: time-series weight. =0.5, spatial weight =0.3, type weight =0.2, weight constraint formula =1, substitute the three dimension parameters of the multidimensional difference vector into the weighted summation rule, and calculate the weighted comprehensive difference value of the two event vectors. The calculation formula is: , A larger value indicates a greater overall difference between the two event vectors; a smaller value indicates a smaller overall difference. This applies to all combinations of two event vectors within the current time-series sliding window. Perform a full traversal and extract the maximum value, which is then denoted as the maximum weighted difference value. This value serves as the upper limit benchmark for the degree of difference within the current window; the weighted composite difference value is mapped to the standard numerical range of [0, 1] through linear normalization to obtain the data correlation strength of the dual event vectors. The calculation formula is: The numerical meaning of the data association strength S is strictly defined. The closer to 1, the higher the temporal, spatial, and typological correlation between the two event vectors, and the more likely they are consecutive real defective products or jitter signals triggered in the same instance. The closer to 0, the lower the correlation between the two event vectors across all dimensions, indicating that they are independent and unrelated defect-triggered events.

[0025] Step 224: Extract the data correlation strength of each event vector, and construct a dynamic coupled field model with the event vector as the mass point and the correlation strength as the coupling coefficient, to obtain the set of coupled field parameters describing the virtual mechanical relationship between the event vectors. Specifically, the front-end data acquisition device performs a full traversal of all event vector combinations that have been calculated within the time-series sliding window, completely extracting the data correlation strength values ​​between each pair of events, and using this as the core basis to construct the dynamic coupled field model. This model transforms the discrete defect trigger signal of the production line into a mass point coupling system that conforms to the laws of physics and mechanics, and equates the data fluctuations caused by operation jitter and accidental touches to the abnormal motion of the mass points. The entire model construction process is without simplification or omission. Specifically, each initial event vector is equivalent to an independent rigid mass point in the dynamic coupled field, denoted as... Where k is the unique number of the particle, ranging from 1, 2, 3, ..., m, and m represents the total number of initial event vectors within the current time-series sliding window. Each particle carries fixed physical attributes, all derived from the normalized fields of the initial event vectors. The time attribute is the millisecond-level timestamp of the event vector. The spatial attribute is the three-dimensional coordinates after elliptical axis alignment transformation. The type attribute is the defect type code. The above attributes together constitute the basic state parameters of a particle; any two calculated particles... and Data correlation strength between , defined as the coupling coefficient between two particles in a dynamic coupled field, has its numerical range strictly limited to [0, 1]. The magnitude of the coupling coefficient directly determines the strength of the virtual mechanical interaction between the particles. When the value approaches 1, it indicates a very high correlation between the two particles, corresponding to continuous jitter signals from the same operation in a production scenario or continuous defective products of the same type at the same location; when When the value approaches 0, it indicates that the two particles are unrelated, corresponding to independent defective product triggering events in the production scenario. A three-dimensional coupling boundary is preset around each particle; only particles within the boundary will experience virtual mechanical coupling, avoiding interference from irrelevant events. The time dimension's radius of action is preset to 500 milliseconds, meaning only particles with a timestamp difference of less than 500 milliseconds are coupled. The spatial dimension's radius of action is preset to 500 millimeters, meaning only particles with a three-dimensional Euclidean distance of less than 500 millimeters are coupled. The type dimension is limited to the same defect type code, meaning only... = The particles exhibit strong coupling; by integrating all core parameters of the dynamic coupled field, a standardized coupled field parameter set is formed, which includes a set of particle numbers. Set of spatiotemporal properties of a particle Set of pairwise particle coupling coefficients A set of boundary parameters for coupling effects, which fully describes the virtual mechanical relationships between all event vectors.

[0026] Step 225: Based on the coupling field parameter set, simulate the data fluctuations of each discrete defect trigger signal as the mechanical motion trajectory of the corresponding particle under the action of the dynamic coupling field. Calculate the virtual force on the particle based on the data correlation strength, derive the virtual velocity and virtual acceleration of the particle, and obtain the motion state parameters of each event vector. Specifically, the front-end data acquisition unit, based on the coupling field parameter set, completely simulates the data fluctuations of the discrete defect trigger signal as the forced mechanical motion trajectory of the particle in the dynamic coupling field. The virtual force, virtual acceleration, and virtual velocity of the particle are derived using physical mechanics formulas to form complete motion state parameters. Specifically, the virtual force calculation between particles is performed based on the coupling coefficient for any two particles with coupling interaction. right virtual forces Virtual forces characterize the correlation and traction effect between events, and the calculation formula is as follows: ,in The preset coupling ratio gain coefficient, fixed at 10, is used to map the coupling coefficient to a force value consistent with mechanical calculations. This is used for calculating the virtual force on a single target mass. The net virtual force on a particle is obtained by summing the virtual forces exerted on it by all particles within its coupling range. The calculation formula is: ,in For point mass With coupling range Virtual forces between individual particles; derivation of virtual acceleration of individual particles, with the virtual mass of all particles uniformly set to 1. =1 (dimensionless, simplifies embedded calculations and does not affect mechanical laws), according to Newton's second law of motion. = Derivation of the virtual acceleration of a particle The virtual acceleration directly corresponds to the fluctuation amplitude of the data change rate of the defect trigger signal, and the calculation formula is as follows: The virtual velocity of a particle is calculated using the absolute difference in timestamps between two adjacent particles. The virtual velocity of the particle is calculated by combining virtual acceleration with the time interval. The virtual velocity corresponds to the instantaneous data change rate of the defect trigger signal, and the calculation formula is as follows: The motion state parameter set is fully integrated by standardizing and integrating the timestamp of the mass point, three-dimensional spatial coordinates, defect type code, resultant virtual force, virtual acceleration and virtual velocity to form a motion state parameter set.

[0027] Step 226: Based on the motion state parameters, the virtual acceleration is constrained by a preset upper limit for the acceleration change amplitude to limit the fluctuation range of the data change rate between adjacent time moments, thus obtaining the corrected motion state parameters after acceleration constraint; specifically, this includes: setting a preset upper limit for the acceleration change amplitude based on the actual operating characteristics and jitter issues of the production line. The fixed value is 0.02 events / square milliseconds. This threshold is the maximum allowable value for data change rate fluctuations. Exceeding this value is considered an abnormal fluctuation caused by operational jitter or accidental touch. Virtual acceleration amplitude limiting correction applies to the virtual acceleration of each particle. Perform absolute value determination, and then perform virtual acceleration after adjusting according to preset rules to obtain the amplitude limit. The limiting formula is: ,in For a sign function, when When the value is greater than 0, the value is 1. When the value is less than 0, it is set to -1 to ensure that the direction of acceleration after limiting is consistent with the original direction and does not change the trend of data change. The corrected virtual velocity calculation is based on the corrected virtual acceleration after limiting, and the corrected virtual velocity of the particle is recalculated. Maintain time interval The formula remains unchanged. = The modified motion state parameter set is generated by reintegrating the original spatiotemporal attributes, type attributes, resultant virtual force, modified virtual acceleration, and modified virtual velocity of the mass point to form a modified motion state parameter set. This parameter set has eliminated abnormal fluctuation components, and the data variation range is strictly constrained within a reasonable range.

[0028] Step 227: Based on the corrected motion state parameters, perform debouncing and deduplication processing on the initial event vector, eliminating redundant event vectors generated by multiple triggers of the same defect, and merging temporally adjacent and spatially and type-similar jitter signals into a stable event vector to obtain a deduplicated stable event vector sequence; specifically, this includes: multi-dimensional determination of redundant event vectors, traversing the corrected motion state parameters of all particles, and determining event vectors that satisfy all of the following conditions as redundant event vectors generated by operation jitter or accidental touch, and the coupling coefficient with the previous valid event. >0.9, extremely high correlation; difference with the timestamp of the previous valid event <200 milliseconds, which meets the short-duration jitter triggering characteristics in the background technology; the three-dimensional Euclidean distance to the previous valid event. <100 mm indicates triggering at the same workstation; if the defect type code is completely consistent with the previous valid event, it indicates triggering of the same type of defective product; redundant event vectors are precisely eliminated, permanently removing all event vectors judged as redundant from the initial event vector sequence, retaining only valid event vectors that meet the triggering conditions of real defective products, without losing any real defect data; jitter signal merging processing is performed, for multiple consecutive jitter signals that meet the redundancy judgment conditions, a standardized merging operation is performed, the merging rule is to retain the timestamp of the first triggering event in the group of jitter signals as the timestamp of the merged event, ensuring timing accuracy; the three-dimensional spatial coordinates and defect type code of the original event are retained, without changing the spatial and type attributes of the event; the counting logic is only incremented once to avoid falsely high counts caused by jitter; stable event vector sequence generation, after completing the elimination of redundant events and merging of jitter signals, all remaining event vectors are real valid defect triggering events, rearranged in ascending order of timestamps to form a stable event vector sequence.

[0029] Step 228: Based on the stable event vector sequence, perform frequency domain limiting processing to filter out high-frequency noise components above a preset frequency threshold, retaining the low-frequency principal components that reflect the actual defect occurrence patterns, thus obtaining the limited event vector sequence. Specifically, this includes: through long-term data collection and statistical analysis of similar high-speed assembly production lines, it is determined that the frequency of actual defective product triggering events on the production line is always in the low-frequency range, while interference signals generated by operator hand tremors, electromagnetic interference from production line motors and frequency converters, and data signal transmission glitches are all in the high-frequency range. Using the boundary between the actual defect triggering frequency and the interference signal frequency as the calibration basis, a fixed value for the high-frequency noise cutoff frequency is finally determined. This value is set to 5 Hz. The system uniformly judges all signal components with frequencies higher than this cutoff frequency as high-frequency noise that is not triggered by actual defective products. This type of noise is mainly generated by manual operation tremors, equipment electromagnetic interference, and signal transmission glitches, which will directly cause distortion of defective product count data and needs to be completely filtered out through subsequent processing. First, the stable event vector sequence after deduplication and jitter reduction undergoes time-domain data preprocessing. The timestamps and counts of each event in the sequence are integrated into a continuous time-domain sampled signal. Uniform sampling is performed according to a fixed sampling rate adapted to the industrial embedded system, ensuring the regularity and computability of the signal data. Then, a Fast Fourier Transform (FFT) algorithm optimized for industrial time-series data is activated to transform the preprocessed time-domain sampled signal. This algorithm, without complex redundant calculations, decomposes the continuous time-domain signal into a superposition of multiple basic signals of different frequencies, extracting all frequency components contained within the signal one by one, and calculating the signal amplitude corresponding to each frequency component. Finally, the original time-domain data is completely converted into frequency-domain data. The frequency-domain data allows for a clear distinction between the low-frequency effective components representing the actual occurrence of defective products and the high-frequency noise components representing operational jitter and electromagnetic interference, providing a clear basis for subsequent noise filtering. The system then performs a full-domain traversal of the converted frequency-domain data, sequentially reading the specific value of each frequency component and comparing the value of each frequency component with a preset... The system compares the high-frequency noise cutoff frequency. When the value of a frequency component is higher than the preset high-frequency noise cutoff frequency, the system immediately marks the component as a high-frequency noise component and directly deletes all data of that component, completing the noise removal operation. When the value of a frequency component is less than or equal to the preset high-frequency noise cutoff frequency, the system determines that the component is an effective low-frequency principal component reflecting the actual occurrence pattern of defective products on the production line. The system retains the frequency and amplitude information of the component. The system then starts the inverse fast Fourier transform algorithm matched with the previous fast Fourier transform algorithm to reconstruct the signal of the frequency domain data after filtering out high-frequency noise. Based on the retained effective low-frequency principal components, the algorithm restores the frequency domain data characterized by frequency and amplitude to time domain data characterized by time and event counts. During the restoration process, the system simultaneously performs time-series calibration, data normalization, and amplitude correction on the reconstructed time domain data to ensure that the restored event sequence remains complete and accurate in terms of time order, count values, and event attributes, without data distortion, time sequence disorder, or information loss. After the conversion is completed, an event vector sequence that has undergone frequency domain amplitude limiting processing is generated.

[0030] Step 229: Based on the event vector sequence after amplitude limiting, reorganize it into a smooth and continuous structured event sequence in chronological order. Specifically, this includes: the front-end data acquisition unit performs full-domain temporal calibration and reorganization on the event vector sequence after frequency domain amplitude limiting, using the millisecond-level timestamp of each event vector as the sole sorting criterion to complete full sequence normalization. During the normalization process, the timestamp, three-dimensional spatial coordinates, defect type encoding, and all contextual information of the event vectors are fully preserved, while all invalid data, noise components, and redundant information are removed. After reorganization, a smooth and continuous structured event sequence is finally generated. This sequence has the following characteristics: no operational jitter distortion, no high-frequency noise interference, no redundant counting, temporal continuity, stable data, and fully preserves the core temporal, spatial, and type characteristics of real defective products.

[0031] By using dynamic coupled field modeling, mechanical motion trajectory simulation, acceleration amplitude limiting constraints, multi-dimensional anti-shake deduplication, and frequency domain filtering, the original discrete and easily disturbed defect trigger signals are transformed into smooth and continuous high-quality structured event sequences, thereby improving the authenticity, stability, and reliability of defective product counting data.

[0032] In a preferred embodiment of the present invention, step 3 above involves performing multi-channel aggregation calculations on the structured event sequence to track the flow trajectory of each defect type in the process chain, while simultaneously calculating the instantaneous rate of change and trend acceleration of the cumulative defect amount; and then performing multi-dimensional feature weighted fusion of the instantaneous rate of change, trend acceleration, and the flow trajectory of the defect type to obtain a quality state distribution tensor, which may include: In this embodiment of the invention, step 330 involves decoupling and dividing the sequence into multiple independent defect type data channels based on the type codes attached to each event vector in the structured event sequence. Then, based on the timestamp order and spatial coordinate distribution characteristics of the event vectors within each channel, channel-level time alignment and workstation node aggregation operations are performed to obtain a multi-channel time-series aggregated dataset. Specifically, this includes: using the defect type code value as the sole classification criterion, performing a decoupling and splitting operation on the overall continuous structured event sequence. During decoupling, the code value of each event vector is determined one by one, and all event vectors with the same code are grouped into the same group. Each independent group corresponds to a dedicated defect type data channel, and the final number of channels generated is completely consistent with the total number of defective product types preset on the production line. All independent defect type data channels... The system employs a logical isolation mechanism, ensuring that event data within different channels belongs solely to their corresponding defect type. This eliminates data crosstalk, type overlap, and event confusion. After decoupling and partitioning multiple channels, the system performs channel-level timing alignment for each independent defect type data channel. During this process, the system first extracts the millisecond-level timestamp field of all event vectors within the current channel. Using the timestamp value as the sole criterion, a stable sorting rule is adopted to regularize all event vectors within the channel. The smaller the timestamp value, the earlier it is arranged, and the larger the value, the later it is arranged. Events with identical timestamps retain the original hardware acquisition and reception order. The timing alignment process fully preserves all inherent attributes of each event vector, such as its three-dimensional spatial coordinates and defect type code, and only calibrates the timing position, ensuring that the event vectors within a single channel maintain a strict chronological order.

[0033] During the production line deployment phase, the system pre-configures and stores the spatial coordinate mapping relationships of workstations. The configuration process is based on the physical layout of the production line, setting a unified three-dimensional spatial coordinate reference system with millimeters as the coordinate unit. The reference point at the beginning of the production line is used as the coordinate origin, defining the length direction as the X-axis, the width direction as the Y-axis, and the height direction as the Z-axis. Through precision mapping tools, the spatial coordinates of each independent process workstation on the production line are collected, clarifying the coverage of each process workstation in the three-dimensional coordinate system. That is, each workstation corresponds to a set of X-axis coordinate intervals, Y-axis coordinate intervals, and Z-axis coordinate intervals. The coordinate intervals of all workstations are independent of each other, with no overlap or omission, ensuring that the spatial coordinates of any event vector can only belong to a unique process workstation. At the same time, the number, workstation name, and process link of each process workstation are bound to the corresponding three-dimensional coordinate intervals to form a complete workstation spatial coordinate mapping relationship table.

[0034] The system calls a pre-stored workstation spatial coordinate mapping table. For event vectors that have completed time-series alignment within each independent defect type data channel, it performs workstation node aggregation operations. During the operation, the three-dimensional spatial coordinate values ​​of each event vector are extracted one by one. The X-axis, Y-axis, and Z-axis components of the coordinates are matched and determined dimension by dimension with the three-dimensional coordinate intervals of each process workstation in the mapping table. When all three-dimensional coordinate components of an event vector fall within the coordinate interval of a certain process workstation, the event vector is determined to belong to that process workstation node. After completing the workstation assignment determination for all event vectors, the system classifies them according to process workstation nodes and centrally aggregates and integrates event vectors belonging to the same workstation within the same channel, forming event subsets with workstation nodes as units. At the same time, the event vectors within each workstation node subset are secondarily regularized according to the time-series alignment order to ensure that the event time sequence and assignment are clear and definite within the same workstation. After the full-process standardization process of defect type decoupling, channel-level time-series alignment calibration, and workstation node aggregation operations, the system uniformly aggregates and integrates the regularized data within all independent defect type data channels to form a standardized multi-channel time-series aggregation dataset.

[0035] Step 331: Based on the multi-channel time-series aggregated dataset, extract the event vectors continuously distributed along the time axis within each independent defect type data channel. Construct a node connection mapping based on the upstream and downstream physical relationships of the processes corresponding to the spatial coordinates. Continuously connect the trigger points across workstations to form directional paths, obtaining the flow trajectory of each defect type in the process chain. Specifically, this includes: the system sequentially retrieves the regularized data of each independent defect type data channel from the multi-channel time-series aggregated dataset, extracts all event vectors continuously distributed along the time axis within the channel, and these event vectors correspond to the trigger records of the same defect type at different process workstations on the production line. The system calls the preset data from the production line deployment phase. This table represents the physical upstream and downstream relationships of the production process. During the production line deployment phase, this table is pre-configured based on the complete processing and assembly process of the product. During pre-configuration, all workstations on the production line are sequentially numbered according to the actual physical flow path of the product from raw material input to finished product output. The order of each workstation in the production process is clearly defined, with the processing and inspection workstations the product passes through first defined as upstream workstation nodes, and the processing and inspection workstations the product passes through last defined as downstream workstation nodes. Adjacent workstations form a unique one-way upstream and downstream relationship. The table fully records the workstation number, workstation name, associated process step, directly upstream workstation number, and directly downstream workstation number for each workstation. The system clearly marks the progressive flow relationships between workstations, eliminating loops, reverse flows, and other associations that do not conform to actual production logic. It ensures that the upstream and downstream relationships in the table completely match the physical paths of actual product production, processing, and transportation, with no reversed order, incorrect associations, or missing nodes. This table maintains a fixed configuration during system operation and is only recalibrated and updated when the production line process flow is adjusted. Based on the preset upstream and downstream relationships, the system constructs standardized node connection mappings for all workstations on the production line. Each workstation is treated as an independent basic node, and the physical flow direction of product production is the unique connection direction, establishing unidirectional, loop-free node associations. To ensure that node connections conform to actual production logic, for consecutive trigger events of the same defect type, the system combines the order of event vector timestamps with the directional direction of node connection mapping to sequentially connect defect trigger points across different workstations in a directional manner. This connects the independent defect trigger events scattered across various workstation nodes into a complete continuous directional path. This path fully records the entire process of the same defect type being transmitted and transferred from upstream workstations to downstream workstations, which is the flow trajectory of the defect type in the production line process chain. The system performs the same processing flow for all defect type data channels, ultimately generating the full process flow trajectory corresponding to all defect types.

[0036] Step 332: Based on the event trigger timestamps corresponding to each process node in the flow trajectory, the triggered events are cumulatively counted within a preset sliding statistical period to obtain the time series of cumulative defects. The quotient of the incremental difference of the cumulative defects within adjacent sliding statistical periods and the period duration interval is calculated to obtain the instantaneous rate of change of the cumulative defects. Specifically, this includes: the system first completes the preset configuration of the sliding statistical period, which is a fixed time interval used for counting the number of defects per unit time. The duration is uniformly preset according to the production line's production rhythm and is the standard time unit for counting the number of defects; the system classifies each defect trigger event into a specific time unit based on the event trigger timestamps corresponding to each process station node in the flow trajectory of each defect type. Within the corresponding sliding statistical period, the number of defect triggering events is counted cumulatively for each period. The accumulated defect counts are arranged sequentially according to the sliding statistical period, forming a complete defect accumulation time series. This series directly reflects the continuous accumulation of defect counts over time. Based on the defect accumulation time series, the system extracts the defect accumulation counts from two adjacent sliding statistical periods, calculates the incremental difference between the two periods, and then divides this incremental difference by a fixed interval of the sliding statistical period. The final result is the instantaneous rate of change of the defect accumulation, which characterizes how quickly the defect accumulation count changes per unit time. The calculation formula is... ,in The instantaneous rate of change of the cumulative defect quantity is used to quantify the magnitude of change in the cumulative defect quantity per unit time. For the first The total number of defects obtained within each sliding statistical period; For the first The system calculates the total number of defects within each sliding statistical cycle. The system performs the above statistical and calculation operations on each defect type and each workstation, and finally generates full-dimensional, full-coverage instantaneous change rate data of the cumulative defect amount.

[0037] Step 333: Based on the numerical evolution sequence of the instantaneous rate of change of the cumulative defect quantity over a continuous sliding statistical cycle, calculate the quotient of the gradient difference between adjacent instantaneous rates of change and the corresponding cycle interval, and convert the gradient difference into a trend acceleration characterizing the direction and speed of the defect accumulation trend. Specifically, this includes: for each independent defect type, the system arranges the calculated instantaneous rate of change of the cumulative defect quantity according to the time sequence of the preset sliding statistical cycle. All instantaneous rate of change values ​​are progressively arranged in a continuous and complete numerical evolution sequence of instantaneous rate of change. This sequence, with the sliding statistical cycle as the horizontal axis of time and the instantaneous rate of change as the vertical axis of value, completely and continuously reflects the dynamic trend of the cumulative rate of the defect type as production time progresses. It can intuitively and clearly reflect the speed and fluctuation of the increase or decrease in the number of defects. The system performs cycle-by-cycle traversal extraction on the generated numerical evolution sequence of instantaneous rate of change, sequentially selecting any two adjacent sliding statistical cycles in the sequence, and accurately reading the instantaneous rate of change of the cumulative defect quantity corresponding to the i-th sliding statistical cycle and the instantaneous rate of change of the cumulative defect quantity corresponding to the (i+1)-th sliding statistical cycle. The rate of change is calculated by subtracting the value of the previous period from the value of the subsequent period, yielding the gradient difference between two adjacent instantaneous rates of change. This gradient difference directly reflects the magnitude of the increase or decrease in the defect accumulation rate within adjacent periods. The system then divides the calculated gradient difference by a fixed interval of the sliding statistical period preset during the production line deployment phase, converting the rate change within a unit period into the rate change magnitude within a unit time. The final result is standardized and converted into the trend acceleration of the defect accumulation. This trend acceleration is the core characteristic parameter for quantifying the evolution of the defect accumulation trend, characterizing the direction and speed of change. When the trend acceleration is positive, it indicates a continuous increase in the defect accumulation rate, with the rate of increase in the number of defects accelerating. When the trend acceleration is negative, it indicates a continuous decrease in the defect accumulation rate, with the rate of increase in the number of defects slowing down. When the trend acceleration is zero, it indicates that the defect accumulation rate remains stable with no significant change. The larger the absolute value of the trend acceleration, the more drastic the change in the defect accumulation trend, and the earlier it can reflect abnormal fluctuations in the production line's quality status. The calculation formula is... ,in It is the trend acceleration of the cumulative defect amount, used to quantify the intensity, direction and fluctuation of the defect accumulation trend, and is the core indicator for judging the evolution trend of the production line quality status. The first in the time sequence The instantaneous rate of change of the cumulative amount of defects calculated for the corresponding defect type within a sliding statistical period; The first in the time sequence The instantaneous rate of change of the cumulative amount of defects calculated for the corresponding defect type within a sliding statistical period.

[0038] Step 334: Extract the spatial topological order and physical connection relationships of each process node in the flow trajectory, construct a node coupling weight set representing the dependence of defect propagation in the process chain, and establish a multi-dimensional feature space alignment basis based on the node coupling weight set; specifically, this includes: using the flow trajectory of each defect type in the process chain as the only input data, first performing a global analysis on the flow trajectory, and extracting the spatial topological order and physical connection relationships of each process node in the trajectory one by one. The spatial topological order is used to mark the fixed arrangement position and hierarchical order of each process node in the three-dimensional space of the production line, and the physical connection relationship is used to clarify the unidirectional upstream and downstream association of product flow between process nodes; based on the extraction Based on the spatial topological order and physical connectivity, the system begins constructing a set of node coupling weights to characterize the propagation dependence of defects in the process chain. This construction process consists of three stages: The first stage is process node baseline calibration, where the process node where the defect first appears in the flow trajectory is defined as the source baseline node, and the immediately following process node directly connected to the source node is defined as a first-level propagation node, extending outwards to define the propagation hierarchy. The second stage is node coupling weight calculation. For any pair of upstream and downstream process nodes with a defect propagation relationship, the coupling weight is calculated based on the actual propagation frequency of the defect between the two nodes and the topological hierarchy interval between processes. The calculation formula is as follows: ,in For upstream process nodes With downstream process nodes The node coupling weights between the nodes directly reflect the degree of defect propagation dependence between the two nodes. For defects from upstream nodes within the current statistical period Transmitted to downstream nodes Total number of times; upstream node With downstream nodes The third stage is the collection and integration of weight sets. All coupling weights between process nodes are uniformly organized according to the path order of the flow trajectory to form a node coupling weight set covering the entire process and the entire transmission path. There are no missing data, no duplicate calculations, and no correlation errors in the weight set. After the node coupling weight set is completed, the system uses this weight set as the core basis to establish a multi-dimensional feature space alignment basis that is completely matched with the data flow processing dimension. This basis is a three-dimensional orthogonal structure. The three orthogonal directions correspond to the time evolution dimension, spatial process dimension, and defect type dimension, respectively. Each coordinate point of the basis corresponds one-to-one with the sliding statistical period, process node, and defect type. It can achieve misalignment-free alignment of three types of features: instantaneous change rate, trend acceleration, and defect flow trajectory. Finally, the multi-dimensional feature space alignment basis is output.

[0039] Step 335: Based on the multidimensional feature space alignment basis, perform temporal phase synchronization and dimensional normalization processing on the instantaneous change rate and trend acceleration to obtain a standardized temporal feature sequence. Specifically, this includes: using the multidimensional feature space alignment basis as the processing benchmark, simultaneously calling two types of core temporal feature data—the instantaneous change rate of the cumulative defect amount and the trend acceleration of the cumulative defect amount—and sequentially performing temporal phase synchronization and dimensional normalization processing. First, the temporal phase synchronization processing is performed. The system forcibly calibrates the sliding statistical period time axis corresponding to the instantaneous change rate and trend acceleration with the time evolution dimension of the multidimensional feature space alignment basis to ensure that the sliding statistical data of the same sequence number are synchronized. Within the period, the time scales of the instantaneous rate of change and the trend acceleration completely overlap, eliminating the offset and misalignment of the two types of features in the time dimension, achieving complete synchronization of the time phase, and performing dimensional normalization processing. Since the original dimensions and numerical ranges of the instantaneous rate of change and the trend acceleration are different, they cannot be directly merged. The system uses the extreme value normalization method to perform dimensionless standardization transformation on the two types of features, uniformly mapping all feature values ​​to the standard interval of 0 to 1. After completing the time phase synchronization and dimensional normalization, the system integrates and arranges the processed instantaneous rate of change and trend acceleration in the order of the sliding statistical period to generate a standardized time series feature sequence.

[0040] Step 336: Based on the standardized temporal feature sequence and node coupling weight set, calculate the dynamic cross-correlation degree between each feature channel to obtain a multi-dimensional feature weighted mapping matrix. Then, perform a nonlinear projection operation on the standardized temporal feature sequence based on the multi-dimensional feature weighted mapping matrix to obtain the channel fusion feature flow. Specifically, this includes: calculating the dynamic cross-correlation degree between each feature channel. The feature channels include the standardized instantaneous rate of change channel, the standardized trend acceleration channel, and the node coupling weight channel. The dynamic cross-correlation degree is used to quantify the coupling matching degree between any two feature channels. The higher the correlation degree, the stronger the correlation between the two types of features, and the higher the weight should be during fusion. The specific calculation process is as follows: feature data extraction... From the standardized time-series feature sequence, the standardized instantaneous rate of change and standardized trend acceleration corresponding to each sliding statistical period, each process node, and each defect type are extracted. From the node coupling weight set, the node coupling weights of corresponding coordinate points (same period, same process node, same defect type) are extracted to ensure that the coordinate points of the three types of feature data are completely corresponding and there is no data misalignment. Dynamic cross-correlation is calculated using the Pearson correlation coefficient method, separately calculating the dynamic cross-correlation between the standardized instantaneous rate of change and node coupling weights, the standardized trend acceleration and node coupling weights, the standardized instantaneous rate of change and standardized trend acceleration, and the three sets of features. The calculation formula is unified as follows: ,in The value represents the dynamic cross-correlation degree between two types of features. It is dimensionless and ranges from [0, 1]. The closer the value is to 1, the higher the coupling and matching degree of the two types of features. The closer the value is to 0, the weaker the correlation between the two types of features. When fusing, the weight of this type of feature can be weakened. For the first The standard value of the first type of feature within a sliding statistical period; For the first The standard value of the second type of feature within a sliding statistical period; The first type of feature is the overall numerical mean over all moving statistical periods, dimensionless, and passes through all... The sum divided by the total number of statistical periods get; The second type of feature is the overall numerical mean over all moving statistical periods, dimensionless, and passes through all... The sum divided by the total number of statistical periods We obtain; after calculating the dynamic cross-correlation degree among the three sets of features, we obtain three correlation degree values, which are denoted as follows: (The correlation between instantaneous rate of change and node coupling weight) (The correlation between trend acceleration and node coupling weight) (The correlation between instantaneous rate of change and trend acceleration), these three correlation values ​​will serve as the core weighting coefficients for the subsequent construction of the multidimensional feature weighted mapping matrix.

[0041] A multidimensional feature weighted mapping matrix is ​​constructed. The core function of this matrix is ​​to transform the correlation of three types of features (standardized instantaneous rate of change, standardized trend acceleration, and node coupling weight) into fusion weights, achieving weighted fusion of the three types of features. The specific construction process is as follows: Matrix dimensions are determined by aligning the multidimensional feature space with the three dimensions of the basis (time evolution, spatial process, and defect type). The matrix is ​​a three-dimensional matrix, with rows corresponding to the three feature channels (standardized instantaneous rate of change, standardized trend acceleration, and node coupling weight) and columns corresponding to the three core dimensions of the basis (number of periods in the time evolution dimension, number of nodes in the spatial process dimension, and number of types in the defect type dimension). Each element of the matrix corresponds to a weighting coefficient for a feature channel and a basis coordinate point. Matrix element filling involves using the calculated three sets of dynamic cross-correlation degrees as core weighting coefficients to fill the corresponding elements of the matrix. The filling rule is as follows: the matrix element corresponding to the standardized instantaneous rate of change is filled with... and The average value, i.e. This ensures that the weight of this feature simultaneously considers its correlation with node coupling weights and trend acceleration; the matrix elements corresponding to trend acceleration are standardized and filled. and The average value, i.e. This ensures that the weight of this feature simultaneously considers its correlation with node coupling weights and instantaneous change rate; the matrix elements corresponding to the node coupling weights are filled with... and The average value, i.e. This ensures that the weight of the feature takes into account its correlation with both time-series features. After filling, all elements in the matrix are normalized to ensure that the sum of the weights of the three types of features corresponding to each basis coordinate point is 1, thus avoiding weight imbalance and ultimately forming a complete multidimensional feature weighted mapping matrix.

[0042] After constructing the multidimensional feature weighted mapping matrix, a nonlinear projection operation is performed to map the standardized temporal feature sequence to a unified fusion space, generating a channel fusion feature stream. The specific operations are as follows: Feature projection matching: The standardized temporal feature sequence is matched one by one to the corresponding position of the multidimensional feature weighted mapping matrix according to the coordinate points, ensuring that each set of standardized temporal features can match the corresponding element (weighting coefficient) of the matrix. Nonlinear projection operation: For each coordinate point, the standardized temporal feature is multiplied by the weighting coefficient of the corresponding position in the matrix, and then a nonlinear fusion operation is performed. During the operation, the Sigmoid activation function is used to map the fused feature values ​​to the [0, 1] interval, while retaining the effective feature information with high correlation, such as feature combinations with a correlation close to 1, and weakening the redundant interference information with low correlation, such as feature combinations with a correlation close to 0, to avoid redundant information affecting the subsequent tensor construction. Feature stream integration: The fused feature values ​​of all coordinate points are integrated sequentially according to the three dimensions (time evolution, spatial process, defect type) of the multidimensional feature space aligned with the basis, forming a continuous and regular data stream, which is the channel fusion feature stream.

[0043] Step 337: Based on the channel fusion feature flow, perform tensor skeleton expansion along the time evolution dimension, spatial process dimension, and defect type dimension. Perform feature field tensor product operation and orthogonal basis convergence processing to map the channel fusion feature flow into a unified three-dimensional state space, obtaining the quality state distribution tensor. Specifically, this includes: performing a three-dimensional tensor skeleton expansion operation, specifically as follows: skeleton dimension matching, strictly aligning the three orthogonal dimensions of the basis (time evolution dimension, spatial process dimension, and defect type dimension) according to the multi-dimensional feature space, constructing a three-dimensional tensor skeleton. The three dimensions of the skeleton correspond one-to-one with the three dimensions of the basis, ensuring that the coordinate points of the skeleton are completely consistent with the coordinate points of the basis, with no dimension. The system is designed to eliminate missing coordinates and misalignments. The skeleton structure is calibrated by uniquely calibrating each coordinate point of the 3D tensor skeleton, following the same calibration rules as the base coordinate points. Specifically, coordinate point (i, j, k) corresponds to the i-th sliding statistical cycle, the j-th process node, and the k-th defect type. A unique identifier is assigned to each coordinate point to ensure no confusion or omissions during subsequent feature value filling. The skeleton integrity is verified after construction, ensuring it covers all sliding statistical cycles, all process nodes, and all defect types, with no missing coordinates or structural distortions. Upon successful verification, the tensor skeleton is fully unfolded to obtain a well-structured 3D tensor skeleton.

[0044] The core of this operation is to fill the corresponding coordinate points of the 3D tensor skeleton with the feature values ​​from the channel-fused feature stream, achieving deep fusion of feature information and tensor space. Specifically, the operation involves: feature value matching, matching each feature data point in the channel-fused feature stream to its corresponding coordinate point in the 3D tensor skeleton according to its time period, process node, and defect type, ensuring that each feature data point can find a unique corresponding tensor coordinate point with no matching errors or data omissions; and tensor product operation execution, where for each coordinate point, the feature values ​​from the channel-fused feature stream are... The basis values ​​of the coordinate point in the basis aligned with the multidimensional feature space. Perform a tensor product operation, and use the result as the tensor characteristic value of that coordinate point. The calculation formula is as follows: ,in After the tensor product operation, the tensor feature values ​​corresponding to the i-th sliding statistical cycle, the j-th process node, and the k-th defect type are dimensionless and range from [0, 1]. The larger the value, the worse the production quality status corresponding to the coordinate point (more accumulated defects, faster changes, and stronger transmission dependence). In the channel fusion feature flow, the fusion feature value corresponding to the i-th cycle, j-th process node, and k-th defect type is dimensionless and ranges from [0, 1]. It integrates the information of three core features. For the basis in the multidimensional feature space alignment, the basis value corresponding to the i-th cycle, j-th process node, and k-th defect type is dimensionless and has a value range of [0, 1]. It is used to calibrate the position of the feature value in the tensor space to ensure that the feature value matches the basis.

[0045] After completing the tensor product operation for all coordinate points one by one, preliminary three-dimensional tensor data is obtained. This data has integrated all core feature information, but there are still some feature redundancy and dimensional interference. Therefore, orthogonal basis convergence processing needs to be performed. The specific operation of orthogonal basis convergence processing is as follows: using the multidimensional feature space aligned basis as the orthogonal reference, the preliminary three-dimensional tensor data is converged and calibrated. The least squares method is used to remove redundant feature information in the tensor data that is inconsistent with the orthogonal direction of the basis, and to correct the coordinate point values ​​with large fluctuations, so that the tensor values ​​of all coordinate points converge to a stable state, ensuring the purity and stability of the tensor data. During the convergence process, effective feature information is strictly retained, and no core data related to production quality is lost. The convergence judgment criterion is that the fluctuation range of the values ​​of all coordinate points in the tensor data is less than the preset threshold, the preset threshold is 0.01, and all values ​​fall within the interval [0, 1] without abnormal fluctuations. After the convergence processing is completed, the quality state distribution tensor is finally generated.

[0046] By progressively weighting and fusing defect flow trajectories, instantaneous change rates, and trend accelerations, and constructing tensors, discrete and scattered multi-dimensional quality characteristics are integrated into a unified quality state distribution tensor, thereby enhancing the data analysis capabilities and quality control value of defect counters.

[0047] In a preferred embodiment of the present invention, step 4 above involves performing phase boundary deduction on the quality state distribution tensor to determine the warning phase region of the current quality state, correcting the preset static critical threshold based on the warning phase region to obtain a dynamic critical threshold, and mapping the dynamic critical threshold to form a real-time quality topology dashboard and a hierarchical warning control flow to achieve streaming monitoring and dynamic hierarchical response of defective product count data on the production line. This step may include: In this embodiment of the invention, step 440 involves extracting multidimensional state feature components from the quality state distribution tensor, performing gradient field reconstruction operations along the time evolution dimension and the spatial process dimension to obtain a phase gradient distribution matrix characterizing the evolution trend of the quality state. Specifically, this includes: firstly, performing a global analysis on the input quality state feature tensor, identifying each tensor element corresponding to each sliding statistical period, process node, and defect type, and extracting two types of core features to ensure no omissions or confusion: time dimension features, representing the quality state change trend of the corresponding process node and defect type within each sliding statistical period, such as the quality value fluctuation of a defect in a certain process within a certain period; and spatial dimension features, representing the quality state distribution of each process node and each defect type, such as the concentration range of quality values ​​for a certain type of defect in a certain process node. The extracted two types of features are then standardized, normalizing all feature values ​​to the [0, 1] interval, and gradient calculation is performed, calculating gradients for both the time and spatial dimensions to construct the phase gradient matrix. The time dimension gradient is calculated for the quality state change gradient of the same process node and the same defect type within each sliding statistical period, using the formula: ,in This is the quality status value for the next sliding cycle. Given the current cycle's quality state value and spatial gradient, calculate the gradient of quality state change for the same defect type across different process nodes within the same sliding cycle. The formula is: ,in This represents the quality status value of the adjacent downstream process. This represents the quality status value of the current process. The topological interval between adjacent processes (dimensionless) reflects the quality transmission changes in the spatial dimension. The time dimension gradient and the spatial dimension gradient are integrated and constructed into a four-dimensional phase gradient distribution matrix according to the coordinate order of sliding statistical period-process node-defect type. Each element in the matrix corresponds to a unique time-space-defect combination, clearly presenting the quality change gradient in each dimension.

[0048] Step 441: Extract gradient mutation trajectories from the phase gradient distribution matrix, construct phase separation hypersurfaces along the gradient mutation trajectories, and use the phase separation hypersurfaces to divide the phase boundaries of the mass states, obtaining the phase boundary coordinate set; specifically including: gradient mutation identification, traversing all gradient values ​​in the phase gradient distribution matrix, including temporal and spatial gradients, and using a preset gradient mutation threshold. (Dimensionless, value range [0.1, 0.3], calibrated by production line quality control standards) is used as the criterion. Each gradient value is judged one by one. If the absolute value of the gradient value is greater than... If the gradient value is less than or equal to the gradient abrupt change point, the corresponding coordinate point is determined to be a gradient abrupt change point, and the quality state corresponding to this point is marked as abruptly changed. The complete coordinates (sliding cycle, process node, defect type) and gradient value of this point are recorded simultaneously. If a point is identified as a normal gradient point with no significant abrupt change in quality state, it is not marked. For abrupt change trajectory integration, all marked gradient abrupt change points are connected in series according to the sliding statistical cycle sequence and the upstream / downstream sequence of process nodes to form a complete gradient abrupt change trajectory. Each trajectory corresponds to a defect type and a quality abrupt change path for a process node, ensuring that the trajectory has no breaks or redundancy (repeated abrupt change points are removed). For phase separation hypersurface construction, using the gradient abrupt change trajectory as the core framework and combining it with a multi-dimensional feature space alignment base, a three-dimensional phase separation hypersurface is constructed. The three dimensions of the hypersurface completely correspond to the temporal evolution dimension, spatial process dimension, and defect type dimension of the alignment base. To ensure the hypersurface covers all quality state regions, the curvature of the hypersurface is adjusted during construction to ensure that one side of the hypersurface represents a normal quality region (without abrupt changes) and the other side represents an abnormal quality region (containing all abrupt changes), with clear boundaries and no intersections. The core of phase boundary extraction is to calculate the intersection points of the phase-separating hypersurface with the denoised and corrected quality state feature tensor. All valid intersection points are the boundary points of different quality phases. After coordinate processing, a phase boundary coordinate set is formed. The specific calculation steps are as follows: Based on the constructed three-dimensional phase-separating hypersurface, combined with the multi-dimensional feature space aligned basis, the hypersurface equation is established. Let the three dimensions of the aligned basis be the time evolution dimension. (Corresponding sliding cycle), spatial process dimension (Corresponding to process nodes) and defect type dimension (corresponding to defect type codes), the hypersurface equation can be expressed as: ( , , )=0, where The function is a three-dimensional continuous function of the hypersurface, obtained by fitting the coordinate data of the gradient mutation trajectory. The fitting process uses the least squares method to ensure that the function can accurately fit the trajectory skeleton formed by all gradient mutation points and satisfy the boundary constraint that one side is a normal region and the other side is an abnormal region; the denoised and corrected mass state feature tensor is denoted as... ,in Let i be the i-th sliding cycle (i=1,2,…,n). Let j be the j-th process node (j=1,2,…,m). Let k be the kth defect type (k=1,2,…,p). For the m-th quality feature parameter (m=1,2,…,q), since the phase boundary is only related to the spatial coordinates (t,s,d), it is necessary to extract the coordinate dimension information of the feature tensor to obtain the coordinate set of the feature tensor. This set contains all the quality state coordinates after denoising correction.

[0049] The essence of an intersection point is a point that simultaneously satisfies the conditions on both the phase-separating hypersurface and the mass state characteristic tensor coordinate set. Therefore, a system of equations is established to solve it. In this system, the constraint points of the first equation lie on the phase-separated hypersurface, and the constraint points of the second equation are the effective coordinate points of the mass state after denoising correction, avoiding the solution of virtual intersection points without actual physical meaning. A numerical solution method (Newton's iteration method) is used to solve the above equation system. The specific process is as follows: Initialize the solution interval, based on the coordinate range of the gradient mutation trajectory, and determine the solution intervals in the three dimensions of t, s, and d, i.e. ( These are the minimum and maximum values ​​for all sliding cycles, respectively. ( These are the minimum and maximum codes for all process nodes, respectively. ( The minimum and maximum codes for all defect types are used to narrow the solution range and improve efficiency. Initial iteration points within the solution interval are selected (preferably feature tensor coordinate points near gradient mutation trajectories), and substituted into the hypersurface equation. The residuals of the equations are calculated. If the residuals are less than the preset accuracy threshold (dimensionless, value 0.001, calibrated according to the accuracy requirements of quality control), the point is determined to be an intersection point. If the residuals are greater than the accuracy threshold, the iteration points are updated using the Newton-Raphson iteration method, and the residuals are calculated repeatedly until the residuals meet the accuracy requirements or reach the maximum number of iterations (the preset maximum number of iterations is 100 to avoid iteration divergence). All intersection points obtained are screened, and invalid intersection points are removed. Invalid intersection points include points whose coordinates exceed the actual coordinate range of the feature tensor, points whose deviation from the actual coordinates in the mass state feature tensor is too large (deviation value greater than 0.01), and duplicate intersection points. The intersection points retained after screening are the effective boundary points of different mass phases. The coordinates of the screened effective boundary points are adjusted according to the sliding period. Process nodes Defect types The coordinates are sorted in order, and a unified coordinate format is adopted (retaining 6 decimal places to ensure precision). After removing duplicate coordinates, they are organized into a phase boundary coordinate set. , where k is the number of valid boundary points, and the coordinate set needs to be labeled with the gradient value corresponding to each boundary point.

[0050] Step 442: Perform topological projection mapping on the current quality state distribution tensor based on the phase boundary coordinate set, calculate the relative approximation distance and state evolution rate between the projected state points and each phase boundary, and determine the warning phase region where the current quality state is located based on the relative approximation distance and state evolution rate to obtain the warning phase region determination result; specifically including: topological projection operation, projecting the quality state feature tensor onto the phase separation hypersurface to complete topological projection matching. During the projection process, ensure that the quality state value corresponding to each sliding cycle-process node-defect type can be mapped to the corresponding position on the hypersurface without projection deviation; at the same time, record the coordinates of the projected state points to form a projected state set, ensuring that the projected points correspond to the positions of the phase boundary coordinate set; phase region determination, combining the phase boundary coordinate set, perform phase region determination for each projected quality state point. The determination criteria are: if the projected point is located on the normal quality region side of the hypersurface, and the corresponding gradient value does not exceed the normal quality region side of the hypersurface, then the phase region determination is performed. If the projection point is located on the abnormal quality region side of the hypersurface, and the gradient value exceeds the normal phase region, then it is determined to be a normal phase region; However, if the criteria for severe anomaly are not met, it is determined to be a warning phase zone; if the gradient value corresponding to the projection point far exceeds... If the quality status value exceeds the normal range by more than 3 times, it is judged as an emergency phase zone; phase zone judgment is performed on all projection points one by one to form a phase zone judgment result set, clarifying the phase zone type (normal, early warning, emergency) corresponding to each sliding cycle-process node-defect type; phase drift compensation is calculated, and according to the phase zone judgment results, the corresponding drift compensation amount is matched for different phase zones, and the compensation amount for normal phase zones is calculated. Dimensionless, with a value range of [0, 0.05], used to adapt to minor fluctuations in normal quality conditions; early warning phase region compensation amount. Dimensionless, with a value range of [0.05, 0.2], used to correct thresholds and improve anomaly detection sensitivity; emergency phase region compensation amount. Dimensionless, with a value range of [0.2, 0.3], it is used to enhance anomaly identification and adapt to rapid response to serious quality problems. The value of the compensation amount is strictly calibrated according to the production line quality control standards to ensure that it matches the quality status.

[0051] Step 443: Based on the early warning phase zone determination result, the preset phase zone offset mapping rule is invoked to calculate the phase drift compensation amount. The phase drift compensation amount is then algebraically superimposed with the preset static critical threshold to obtain the dynamic critical threshold. Specifically, this includes: dividing all processes on the production line into critical processes and ordinary processes. The static threshold for critical processes is set more strictly (lower values), while the static threshold for ordinary processes is relatively lenient, ensuring stricter quality control for critical processes. Different static thresholds are set according to the severity of defects, with lower static thresholds for severe defects and higher static thresholds for minor defects, avoiding abnormal omissions or misjudgments due to unreasonable threshold settings. The static threshold range for critical processes + severe defects is 0.3-0.5, the static threshold range for critical processes + general defects is 0.4-0.6, and the static threshold range for ordinary processes + minor defects is 0.6-0.7. All static thresholds are dimensionless and strictly follow the threshold rule of critical processes < ordinary processes and severe defects < general defects, ensuring that the thresholds are highly consistent with quality control requirements.

[0052] The system retrieves a preset static threshold mapping table and, based on the current combination of sliding cycle, process node, and defect type, calls the corresponding static critical threshold (denoted as ). To ensure that the threshold called perfectly matches the current processing object and there are no calling errors, dynamic threshold calculation is performed one by one for each sliding cycle-process node-defect type combination to ensure that each combination has a unique corresponding dynamic threshold. Specific details are as follows: data matching, firstly, the phase drift compensation amount (denoted as...) is... (dimensionless) and the corresponding combination of static thresholds ( Matching is performed to ensure that the compensation amount corresponds to the same sliding cycle-process node-defect type combination as the static threshold, without misalignment; algebraic superposition calculations are performed strictly according to the formula. Perform the calculation, where For each combination of dynamic critical thresholds, the calculation results are immediately checked after each combination is completed to confirm there are no calculation errors, such as omissions in superposition or numerical deviations. If errors are found, the calculation is immediately recalculated to ensure that all dynamic thresholds conform to the rule that critical process thresholds < ordinary process thresholds and severe defect thresholds < general defect thresholds. The dynamic thresholds for each sliding cycle-process node-defect type combination are checked one by one, and double verification is performed to ensure no deviation. Interval verification is also performed to confirm all dynamic thresholds. All values ​​are within the reasonable range of [0.3, 1.0]. If the value is below 0.3, it is adjusted to 0.3 (minimum quality control threshold); if the value is above 1.0, it is adjusted to 1.0 (maximum quality anomaly threshold) to avoid subsequent judgment errors due to thresholds exceeding the reasonable range. The matching degree between the dynamic threshold and the phase zone type is checked. The dynamic threshold for emergency phase zones (such as phase zones corresponding to severe defects) must be lower than the thresholds for normal and warning phase zones to ensure higher sensitivity in anomaly identification under emergency conditions. If there is a mismatch, the compensation amount is readjusted. Continue until the matching requirements are met; correct the threshold deviations and calculation errors found during the verification one by one, and record the correction content synchronously to ensure that each dynamic threshold can correspond to the combination of sliding cycle-process node-defect type without misalignment or omission. Organize all the verified and corrected dynamic thresholds in the order of sliding cycle, process node, and defect type to form a standardized set of dynamic thresholds.

[0053] Step 444: Perform topological projection alignment of the dynamic critical threshold and the quality state distribution tensor along the spatial process dimension. Calculate the deviation gradient between the state values ​​of each process node and the dynamic critical threshold, and construct a real-time quality topology mesh representing the topological relationship of quality state transmission between workstations, obtaining the underlying data mapping set of the real-time quality topology dashboard. Specifically, this includes: performing topological projection alignment of the dynamic critical threshold set and the quality state distribution tensor along the spatial process dimension. The specific process is as follows: using the spatial process dimension of the multi-dimensional feature space alignment basis as a benchmark, ensure that the process node dimension of the dynamic critical threshold set and the quality state distribution tensor are aligned, i.e., the dynamic threshold and quality state feature value of the same process node, the same sliding statistical period, and the same defect type correspond one-to-one, without coordinate misalignment. During the alignment process, check the coordinate information of each sliding statistical period-process node-defect type combination to ensure that the dynamic threshold... With quality state characteristic value Matching involves calculating the deviation gradient between the state values ​​of each process node and the dynamic critical threshold. This gradient characterizes the degree of deviation between the quality state and the dynamic threshold. The calculation formula is as follows: ,in Let be the deviation gradient for the i-th sliding statistical cycle, the j-th process node, and the k-th defect type. A positive value indicates that the quality status exceeds the dynamic critical threshold and is in the abnormal range; the larger the value, the more severe the abnormality. A negative value indicates that the quality status is below the dynamic critical threshold and is in the normal range. A real-time quality topology grid is constructed to represent the topological relationship of quality status transmission between workstations. This grid is the core architecture for visually presenting the quality correlation between processes. The construction process is as follows: based on the actual process flow of the production line and the physical upstream and downstream relationship table of processes, each process node is treated as an independent node of the topology grid, with a unique identifier and name for each node. Completely consistent with the process nodes, node attributes include process number, workstation location, defect type adaptation list, etc. The mesh connection relationship is completely consistent with the physical upstream and downstream relationship of the process. Directed lines connect upstream and downstream process nodes, and the connection direction is consistent with the production flow direction and quality transmission direction. Each connection line is labeled with the topological interval and quality transmission weight between processes to characterize the tightness of quality transmission between processes. The transmission weight between core processes is higher to ensure that the mesh can reflect the quality status transmission relationship between workstations. A real-time quality topology dashboard underlying data mapping set is generated. The specific process is as follows: the calculated deviation gradient is... Dynamic critical threshold Quality state characteristic value The early warning phase zone determination results are filled into the corresponding nodes of the real-time quality topology grid one by one. At the same time, auxiliary information such as detailed defect type, quality status change trend, and data timestamp are added to each node to ensure that the information of each grid node is complete and without missing information. All data in the grid is standardized and sorted in the order of sliding statistical cycle-process node-defect type, and the data format is unified and three decimal places are retained to form a standardized underlying data mapping set.

[0054] Step 445 involves performing spatial coordinate matching and threshold exceedance verification on the current defective product count events in the underlying data mapping set of the real-time quality topology dashboard. This identifies abnormal node clusters that exceed dynamic critical thresholds, and calculates the state diffusion potential field based on the spatiotemporal distribution of these abnormal node clusters. A warning level sequence is then formed based on the intensity distribution of the state diffusion potential field, resulting in a tiered warning control flow data frame. Specifically, this includes performing spatial coordinate matching and threshold exceedance verification on the current defective product count events in the underlying data mapping set of the real-time quality topology dashboard. The specific process is as follows: extracting the deviation gradient from the underlying data mapping set. Using a deviation gradient greater than 0 as the criterion, i.e., the quality state exceeds the dynamic critical threshold, all nodes corresponding to abnormal quality states are identified. The complete coordinates (sliding statistical cycle, process node, defect type), deviation gradient value, and warning phase type of each abnormal node are recorded. Spatial coordinate matching is performed on the identified abnormal nodes, grouping those with the same sliding statistical cycle, the same defect type, and adjacent process nodes into an abnormal node cluster. Each abnormal node cluster is labeled with its corresponding defect type, number of abnormal nodes, and core abnormal node (the node with the largest deviation gradient), ensuring clear classification of abnormal nodes and reflecting the extent of defect spread. Based on the spatiotemporal distribution of the abnormal node clusters, a state diffusion potential field is calculated. This potential field characterizes the degree of diffusion risk of the abnormal node cluster, and the calculation formula is as follows: ,in The value represents the state diffusion potential field strength of the abnormal node cluster corresponding to the i-th sliding statistical cycle, the j-th process node, and the k-th defect type. The larger the value, the higher the risk of abnormal diffusion and the higher the warning level. Indicates the current cluster of abnormal nodes. The deviation gradient of the m-th node within the cluster of anomalous nodes is taken from the underlying data mapping set; The process distance between the j-th process node and the m-th node in the cluster is taken from the process physical upstream and downstream relationship table; The preset diffusion attenuation reference value has a range of [1, 3] and is used to calibrate the attenuation rate of the diffusion potential field to avoid deviations in the potential field calculation due to excessive process spacing.

[0055] Based on the intensity distribution of the state diffusion potential field, a warning level sequence is formed, generating hierarchical warning control flow data frames. The specific process is as follows: the system presets four warning levels (level 1 to level 4), corresponding to different potential field intensity ranges. Level 1 warning (minor anomaly) corresponds to... <0.5, Level II warning (moderate abnormality) corresponds to 0.5≤ <1.0, Level 3 warning (severe anomaly) corresponds to 1.0≤ <1.5, Level 4 warning (extreme anomaly) corresponds to ≥1.5, based on the potential field strength of each abnormal node cluster, match the corresponding early warning level to form an early warning level sequence. The sequence contains core information such as the early warning level, abnormal node coordinates, deviation gradient, and diffusion potential field strength of each abnormal node cluster. The early warning level sequence and the detailed information of the abnormal node cluster are encapsulated and a hierarchical early warning control flow data frame is generated according to a standardized data format.

[0056] Step 446: Based on the warning level sequence in the hierarchical early warning control flow data frame, retrieve the preset production line control strategy mapping library to perform strategy matching, and convert the warning level sequence into a dynamic intervention instruction set. Specifically, this includes: parsing the hierarchical early warning control flow data frame to extract core information. The specific process is as follows: parse the standardized format of the data frame, extract the warning level sequence, abnormal node cluster information, defect type, diffusion potential field strength, sliding statistical period, and other core content to clarify the abnormal location, degree, and type of the current production line, and identify the workstations and defect types that require control operations, laying the foundation for control strategy matching. During the parsing process, ensure that the extracted information is complete, without omissions or deviations, and completely consistent with the hierarchical early warning control flow data frame. Retrieve the preset production line control strategy mapping library and perform strategy matching. The specific process is as follows: the production line control strategy mapping library is a set of standardized strategies preset in the system, containing... The system includes control strategies corresponding to four warning levels, with each strategy corresponding to a specific warning level and adaptable to different defect types and process nodes. Specifically: Level 1 warning (minor anomaly) corresponds to an increased inspection frequency strategy, requiring no machine shutdown but only increasing the inspection frequency of the corresponding process to closely monitor changes in quality status; Level 2 warning (moderate anomaly) corresponds to a cycle time adjustment strategy, appropriately reducing the production cycle time of the corresponding process, investigating the cause of the anomaly, and preventing defect spread; Level 3 warning (serious anomaly) corresponds to a partial shutdown strategy, suspending production of the process corresponding to the anomaly cluster, organizing personnel to comprehensively investigate the fault, and promptly eliminating the anomaly; Level 4 warning (extreme anomaly) corresponds to a full-line shutdown strategy, suspending all production processes, thoroughly investigating core faults, and preventing batch defects. Based on the parsed warning level sequence, combined with the process node and defect type of the anomaly cluster, the system retrieves the control strategy mapping library and matches the corresponding control strategy to ensure the strategy's relevance and applicability.

[0057] The process of converting the early warning level sequence into a dynamic intervention instruction set is as follows: the matched control strategy is converted into standardized dynamic intervention instructions. Each instruction contains a clear execution station (the process node corresponding to the abnormal node cluster), operation content (such as inspection encryption, cycle adjustment, shutdown investigation), execution priority (level 4 early warning instructions have the highest priority, level 1 early warning instructions have the lowest priority), response time limit (set according to the early warning level, the higher the early warning level, the shorter the response time limit), execution standards (such as inspection frequency, cycle adjustment range), and other core content to ensure that the instructions can be recognized and executed by the underlying control terminal. All dynamic intervention instructions are sorted according to execution priority and organized into a dynamic intervention instruction set. The instruction set format is adapted to the receiving standard of the underlying control terminal.

[0058] Step 447: Based on the dynamic intervention instruction set, a tiered control signal is sent to the underlying control terminal of the corresponding quality inspection station. Simultaneously, the underlying data mapping set of the real-time quality topology dashboard is triggered to refresh node status and perform visualization rendering, achieving streaming monitoring and dynamic tiered response of the production line defective product count data. Specifically, this includes converting the dynamic intervention instruction set into tiered control signals recognizable by the underlying control terminal. The underlying control terminal (corresponding to each production station and quality inspection station) can only recognize electrical signals of a specific format. Therefore, each instruction in the dynamic intervention instruction set needs to be converted into a corresponding tiered control signal. The signal parameters, such as signal frequency and amplitude, must be consistent with the receiving standard of the underlying terminal to ensure that the terminal can recognize the instruction. The content is free from signal distortion or identification errors. During the conversion process, the execution priority and response time of the instructions are preserved to ensure that the order of signal issuance is consistent with the instruction priority. Graded control signals are issued to the underlying control terminals of the corresponding quality inspection workstations according to the warning priority. The specific process is as follows: according to the execution priority of the dynamic intervention instruction set, the control signals corresponding to the fourth-level warning are issued first, followed by the third, second, and first-level warning signals in sequence, ensuring that emergency anomalies can be responded to quickly. During the issuance process, the issuance time, receiving terminal, and instruction content of each signal are recorded simultaneously, and the signal reception status is monitored in real time. If a terminal fails to receive the signal, it is immediately reissued to ensure that all control signals are delivered to the corresponding workstations without omission or delay.

[0059] The real-time quality topology dashboard's underlying data mapping set is synchronously triggered to refresh node status and perform visualization rendering. The specific process is as follows: Based on the issuance of control signals and the production line's real-time quality data, such as newly added defective product counts and changes in quality status, the relevant data in the underlying data mapping set is refreshed in real time, including deviation gradients, warning phase zones, and quality status characteristic values. At the same time, the nodes of the topology mesh are visualized according to the warning level, using different colors to distinguish different warning levels, such as yellow for level one warning, orange for level two, red for level three, and dark red for level four. The location, spread range, and degree of abnormality of abnormal node clusters are clearly presented, which facilitates managers to grasp the production line's quality status in real time and make quick decisions.

[0060] This system enables streaming monitoring and dynamic tiered response of defective product count data on the production line, forming a closed-loop control system. It continuously collects defective product count data and quality status data from each workstation on the production line, compares them in real-time with dynamic critical thresholds, and monitors the execution effect of control commands. If an anomaly is detected as eliminated (deviation gradient becomes negative), a recovery command is immediately issued to adjust production parameters to normal, and the node rendering status of the topology dashboard is updated synchronously. If an anomaly is detected as continuing to spread (increased diffusion potential field strength, upgraded warning level), the warning level is immediately upgraded, the control strategy mapping library is re-searched, new dynamic intervention commands are generated, and control signals are updated and issued. If a new abnormal node is detected, the above process is repeated, achieving uninterrupted streaming monitoring and dynamic tiered response of defective product count data, ensuring stable production line quality and reducing the generation of defective products.

[0061] By analyzing phase gradients and extrapolating phase boundaries, quality anomaly trends can be identified. Dynamic critical threshold correction can be used to improve the accuracy of anomaly identification. Real-time quality topology grids can be used to visualize quality status. Hierarchical early warning and dynamic intervention can be used to achieve rapid response to anomalies and improve the real-time performance of production line defective product count data monitoring.

[0062] like Figure 2 As shown, embodiments of the present invention also provide a production line defective product counter data stream processing system, comprising: The acquisition and identification module is used to acquire discrete defect triggering signals triggered by each quality inspection station on the production line, and add multi-dimensional context identifiers including timestamps, spatial coordinates and type codes to the discrete defect triggering signals to obtain an initial event vector. The dynamic coupling module is used to construct a time-series sliding window based on the initial event vectors and calculate the data correlation strength between each event vector within the time-series sliding window. Based on the data correlation strength, a dynamic coupling field model is constructed to simulate the data fluctuations of discrete defect trigger signals as mechanical motion trajectories. By limiting the acceleration change amplitude of the data change rate, anti-shake deduplication and frequency domain limiting processing are performed on the initial event vectors to obtain a smooth and continuous structured event sequence. The aggregation calculation module is used to perform multi-channel aggregation calculations on structured event sequences, track the flow trajectory of each defect type in the process chain, and calculate the instantaneous change rate and trend acceleration of the cumulative defect amount. The instantaneous change rate, trend acceleration and the flow trajectory of the defect type are then fused using multi-dimensional feature weighting to obtain the quality state distribution tensor. The dynamic early warning module is used to perform phase boundary deduction on the quality state distribution tensor, determine the early warning phase region of the current quality state, and correct the preset static critical threshold according to the early warning phase region to obtain the dynamic critical threshold. Based on the dynamic critical threshold, a real-time quality topology dashboard and hierarchical early warning control flow are formed to realize the streaming monitoring and dynamic hierarchical response of the defective product count data of the production line.

[0063] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0064] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A production line defective counter data flow processing method, characterized by, The method includes: Step 1: Collect discrete defect triggering signals triggered by each quality inspection station on the production line, add multi-dimensional context identifiers including timestamps, spatial coordinates and type codes to the discrete defect triggering signals, and obtain the initial event vector; Step 2: Construct a time-series sliding window based on the timestamps of the initial event vectors, and calculate the data correlation strength between each event vector within the time-series sliding window; construct a dynamic coupled field model based on the data correlation strength, simulate the data fluctuations of the discrete defect trigger signal as a mechanical motion trajectory, and perform anti-shake deduplication and frequency domain limiting processing on the initial event vectors by limiting the acceleration change amplitude of the data change rate, to obtain a smooth and continuous structured event sequence; Step 3: Perform multi-channel aggregation calculation on the structured event sequence to track the flow trajectory of each defect type in the process chain, and calculate the instantaneous change rate and trend acceleration of the cumulative defect amount; then perform multi-dimensional feature weighted fusion of the instantaneous change rate, trend acceleration and the flow trajectory of the defect type to obtain the quality state distribution tensor. Step 4: Perform phase boundary deduction on the quality state distribution tensor to determine the warning phase region of the current quality state. Correct the preset static critical threshold according to the warning phase region to obtain the dynamic critical threshold. Map the dynamic critical threshold to form a real-time quality topology dashboard and hierarchical warning control flow to realize streaming monitoring and dynamic hierarchical response of defective product count data on the production line.

2. The production line defective counter data flow processing method according to claim 1, characterized by, Collect discrete defect trigger signals from each quality inspection station on the production line, add a multi-dimensional context identifier including timestamp, spatial coordinates, and type code to the discrete defect trigger signals, and obtain an initial event vector, including: Defect detection sensors are deployed at each quality inspection station. When a defective product passes through the inspection area, a discrete defect trigger signal in the form of an electrical pulse is obtained and transmitted to the front-end data acquisition unit to obtain the original pulse sequence. The rising or falling edge of each pulse in the original pulse sequence is captured, and a timestamp corresponding to the time is added to each pulse to obtain a pulse event stream with time sequence identification; Based on the quality inspection station number from which each pulse originates in the pulse event stream, the three-dimensional spatial coordinate value of the station in the physical layout of the production line is retrieved from the preset spatial coordinate mapping table, and the three-dimensional spatial coordinate value is added as a spatial identifier to the corresponding pulse event to obtain a positioning event stream with timestamp and spatial coordinates. Based on the current inspection items set at the quality inspection station corresponding to each pulse event, a preset defect type code is assigned to each event in the location event stream as a type identifier and added to the event to obtain a three-dimensional identified event stream with timestamp, spatial coordinates and type code; Each event in the 3D event stream is encapsulated in chronological order of its timestamp into a structured data tuple containing three fields: timestamp, spatial coordinates, and type code. This is the initial event vector.

3. The production line defective counter data flow processing method according to claim 2, characterized by, Construct a time-series sliding window from the initial event vectors according to their timestamps, and calculate the data correlation strength between each event vector within the time-series sliding window, including: Extract the timestamp attached to each event vector from the initial event vector, and sort all event vectors according to the order of the timestamps to obtain a sequence of event vectors arranged in ascending order of time; Based on preset window length and sliding step parameters, multiple consecutive temporal sliding windows are sequentially divided from the starting position on the event vector sequence arranged in ascending order of time. Each temporal sliding window covers a fixed time span or a fixed number of event vectors, resulting in a windowed event vector grouping sequence. For each time-series sliding window, the event vectors are grouped, and any two event vectors in the group are extracted in turn. The timestamp difference, Euclidean distance between the spatial coordinates and the matching coefficient of the type encoding of the two event vectors are calculated respectively to form a multidimensional difference vector. Based on the multidimensional difference vector, a value between 0 and 1 is calculated according to the preset weighted summation rule, and the value is used as the data correlation strength between the two event vectors.

4. The production line defective counter data flow processing method according to claim 3, characterized by, A dynamic coupled field model is constructed based on the data correlation strength. The data fluctuations of discrete defect trigger signals are simulated as mechanical motion trajectories. By limiting the acceleration variation amplitude of the data change rate, anti-jitter deduplication and frequency domain limiting processing are performed on the initial event vector to obtain a smooth and continuous structured event sequence, including: Extract the data correlation strength of each event vector, construct a dynamic coupled field model with the event vector as the mass point and the correlation strength as the coupling coefficient, and obtain the coupled field parameter set describing the virtual mechanical relationship between the event vectors; Based on the set of coupled field parameters, the data fluctuation of each discrete defect trigger signal is simulated as the mechanical motion trajectory of the corresponding particle under the action of the dynamic coupled field. The virtual force on the particle is calculated based on the data correlation strength, and the virtual velocity and virtual acceleration of the particle are derived to obtain the motion state parameters of each event vector. Based on the motion state parameters, the virtual acceleration is constrained by a preset upper limit of the acceleration change amplitude, limiting the fluctuation of the data change rate between adjacent time moments, and thus obtaining the corrected motion state parameters after acceleration constraint. Based on the corrected motion state parameters, the initial event vector is subjected to anti-shaking and deduplication processing to remove redundant event vectors caused by multiple triggers of the same defect, and jitter signals that are temporally adjacent and spatially and typeally similar are merged into a stable event vector to obtain a deduplicated stable event vector sequence. Based on the stable event vector sequence, frequency domain limiting processing is performed to filter out high-frequency noise components in the sequence that are higher than a preset frequency threshold, while retaining the low-frequency principal components that reflect the actual occurrence pattern of defects, thus obtaining the limited event vector sequence. Based on the event vector sequence after amplitude limiting, it is reorganized into a smooth and continuous structured event sequence in chronological order.

5. The production line defective counter data flow processing method according to claim 4, characterized by, Multi-channel aggregation calculations are performed on structured event sequences to track the flow trajectory of each defect type in the process chain, and the instantaneous rate of change and trend acceleration of the cumulative defect amount are calculated, including: Based on the type encoding attached to each event vector in the structured event sequence, the sequence is decoupled and divided into multiple independent defect type data channels. Based on the timestamp order and spatial coordinate distribution characteristics of the event vectors in each channel, channel-level time sequence alignment and workstation node aggregation operations are performed to obtain a multi-channel time sequence aggregated dataset. Based on a multi-channel time-series aggregated dataset, event vectors continuously distributed along the time axis are extracted from the data channels of each independent defect type. Node connection mapping is constructed according to the upstream and downstream physical relationships of the process corresponding to the spatial coordinates. Continuous trigger points across workstations are sequentially connected to form directional paths, thus obtaining the flow trajectory of each defect type in the process chain. Based on the event trigger timestamps corresponding to each process node in the flow trajectory, the trigger events are accumulated and counted within a preset sliding statistical period to obtain the time series of cumulative defects. The quotient of the incremental difference of cumulative defects within adjacent sliding statistical periods and the period duration interval is calculated to obtain the instantaneous rate of change of cumulative defects. Based on the numerical evolution sequence of the instantaneous rate of change of the cumulative defect amount over a continuous sliding statistical period, the gradient difference between adjacent instantaneous rates of change and the quotient of the corresponding period interval are calculated. The gradient difference is then transformed into a trend acceleration that characterizes the direction and speed of the evolution of the defect accumulation trend.

6. The data stream processing method for a production line defective product counter according to claim 5, characterized in that, The flow trajectory of instantaneous change rate, trend acceleration, and defect type is weighted and fused using multi-dimensional features to obtain the quality state distribution tensor, including: Extract the spatial topological order and physical connection relationship of each process node in the flow trajectory, construct a node coupling weight set to represent the dependence of defect transmission in the process chain, and establish a multi-dimensional feature space alignment basis based on the node coupling weight set; Based on the multidimensional feature space aligned basis, the instantaneous rate of change and trend acceleration are subjected to time-series phase synchronization and dimensional normalization to obtain a standardized time-series feature sequence; Based on the standardized temporal feature sequence and the node coupling weight set, the dynamic cross-correlation degree between each feature channel is calculated to obtain the multi-dimensional feature weighted mapping matrix. Then, a nonlinear projection operation is performed on the standardized temporal feature sequence based on the multi-dimensional feature weighted mapping matrix to obtain the channel fusion feature flow. Based on the channel fusion feature flow, tensor skeleton expansion is performed along the time evolution dimension, spatial process dimension, and defect type dimension. Feature field tensor product operation and orthogonal basis convergence processing are performed to map the channel fusion feature flow into a unified three-dimensional state space to obtain the quality state distribution tensor.

7. The data stream processing method for a production line defective product counter according to claim 6, characterized in that, Phase boundary deduction is performed on the mass state distribution tensor to determine the warning phase region of the current mass state. Based on the warning phase region, the preset static critical threshold is corrected to obtain the dynamic critical threshold, including: Multidimensional state feature components are extracted from the mass state distribution tensor, and gradient field reconstruction operations are performed along the time evolution dimension and the spatial process dimension to obtain the phase gradient distribution matrix that characterizes the evolution trend of the mass state. The gradient mutation trajectory is extracted based on the phase gradient distribution matrix. A phase separation hypersurface is constructed along the gradient mutation trajectory. The phase separation hypersurface is used to divide the phase boundaries of the mass state, and the phase boundary coordinate set is obtained. The mass state distribution tensor at the current moment is topologically projected onto the phase boundary coordinate set. The relative approximation distance between the projected state point and each phase boundary and the state evolution rate are calculated. The warning phase region where the current mass state is located is determined based on the relative approximation distance and the state evolution rate, and the warning phase region determination result is obtained. Based on the early warning phase zone determination result, the preset phase zone offset mapping rule is invoked to calculate the phase drift compensation amount. The phase drift compensation amount is then algebraically superimposed with the preset static critical threshold to obtain the dynamic critical threshold.

8. The production line defective counter data flow processing method according to claim 7, characterized by, Based on dynamic critical threshold mapping, a real-time quality topology dashboard and hierarchical early warning control flow are formed to achieve streaming monitoring and dynamic hierarchical response of defective product count data on the production line, including: The dynamic critical threshold and the quality state distribution tensor are aligned by performing topological projection along the spatial process dimension. The deviation gradient between the state value of each process node and the dynamic critical threshold is calculated. A real-time quality topology grid representing the topological relationship of quality state transmission between workstations is constructed, and the underlying data mapping set of the real-time quality topology dashboard is obtained. Spatial coordinate matching and threshold over-limit verification are performed on the current defective product count events in the underlying data mapping set of the real-time quality topology dashboard. Abnormal node clusters that exceed the dynamic critical threshold are identified, and the state diffusion potential field is calculated based on the spatiotemporal distribution of the abnormal node clusters. The warning level sequence is formed according to the intensity distribution of the state diffusion potential field, and the hierarchical warning control flow data frame is obtained. Based on the warning level sequence in the hierarchical warning control flow data frame, retrieve the preset production line control strategy mapping library to perform strategy matching and convert the warning level sequence into a dynamic intervention instruction set; According to the dynamic intervention instruction set, the underlying control terminal of the corresponding quality inspection station is issued a graded control signal, and the underlying data mapping set of the real-time quality topology dashboard is simultaneously triggered to refresh the node status and visualize the rendering, so as to realize the streaming monitoring and dynamic graded response of the defective product count data of the production line.

9. The production line defective counter dataflow processing method of claim 8, wherein, The dynamic intervention instruction set includes intervention target identifier, intervention intensity parameter, and execution triggering sequence.

10. A production line defective counter data flow processing system implementing the method of any one of claims 1 to 9, characterized by, include: The acquisition and identification module is used to acquire discrete defect triggering signals triggered by each quality inspection station on the production line, and add multi-dimensional context identifiers including timestamps, spatial coordinates and type codes to the discrete defect triggering signals to obtain an initial event vector. The dynamic coupling module is used to construct a time-series sliding window based on the initial event vectors and calculate the data correlation strength between each event vector within the time-series sliding window. Based on the data correlation strength, a dynamic coupling field model is constructed to simulate the data fluctuations of discrete defect trigger signals as mechanical motion trajectories. By limiting the acceleration change amplitude of the data change rate, anti-shake deduplication and frequency domain limiting processing are performed on the initial event vectors to obtain a smooth and continuous structured event sequence. The aggregation calculation module is used to perform multi-channel aggregation calculations on structured event sequences, track the flow trajectory of each defect type in the process chain, and calculate the instantaneous change rate and trend acceleration of the cumulative defect amount. The instantaneous change rate, trend acceleration and the flow trajectory of the defect type are then fused using multi-dimensional feature weighting to obtain the quality state distribution tensor. The dynamic early warning module is used to perform phase boundary deduction on the quality state distribution tensor, determine the early warning phase region of the current quality state, and correct the preset static critical threshold according to the early warning phase region to obtain the dynamic critical threshold. Based on the dynamic critical threshold, a real-time quality topology dashboard and hierarchical early warning control flow are formed to realize the streaming monitoring and dynamic hierarchical response of the defective product count data of the production line.

Images