An industrial equipment edge data cleaning method integrated with an MES

Abstract

This invention relates to the field of electronic digital data processing technology, specifically to a method for edge data cleaning of industrial equipment integrated into a Manufacturing Execution System (MES). The method includes: acquiring the MES production configuration database through an edge computing unit to construct a dynamic data object context dictionary, and encapsulating the original sensor data objects into a standardized data stream; extracting statistical moment features and trend indicators within a time-series sliding window, and identifying noise types by combining dictionary logic constraint thresholds and classification models; dynamically matching and repairing algorithms to perform reconstruction based on the identification results, generating high-fidelity data; mapping the data to the overall equipment efficiency calculation logic of the MES, and iterating the noise classification weight parameters based on feedback deviation closed-loop iteration. This invention, through a business-driven dynamic dictionary and multi-dimensional feature classification and identification logic, achieves self-evolution of the cleaning strategy, significantly improving the processing accuracy and real-time performance of industrial data under complex operating conditions.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, specifically to an edge data cleaning method for industrial equipment integrated into MES. Background Technology

[0002] In distributed industrial computing architectures, edge processing units need to preprocess high-frequency, high-dimensional time-series data generated by sensors. Due to the complex electromagnetic interference and physical equipment wear and tear in industrial environments, raw data objects often contain various noise patterns such as outliers, trend drift, and constant value deadlocks, which directly affect the accuracy of the backend data processing system in determining the equipment operating status and production logic operations.

[0003] Existing data governance solutions typically employ a centralized cloud processing model, transmitting all raw data to a central server via communication links. At the edge, existing cleaning methods are mostly based on single statistical analysis operators, such as using sliding windows to perform mean smoothing or identifying outliers based on standard deviation ranges.

[0004] However, existing technologies have significant limitations: First, the logical thresholds of data cleaning operators are mostly statically configured, lacking a logical connection with the runtime task parameters acquired in real time by the computing system. This leads to the system's inability to identify legitimate data jumps caused by task switching, resulting in misjudgments. Second, existing technologies lack comprehensive logical judgment of multi-dimensional features (such as statistical moments and trend indicators). Under resource-constrained hardware architectures, it is difficult to accurately identify hidden evolutionary trend noise and state dead noise. Finally, the judgment weights of the cleaning algorithm cannot provide closed-loop feedback based on the error results output by the backend logic operation engine, resulting in the data processing accuracy not automatically optimizing as the system runtime increases.

[0005] To address this, a method for cleaning edge data of industrial equipment integrated into MES is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide an industrial equipment edge data cleaning method integrated into MES, which realizes closed-loop self-evolution of multi-dimensional feature noise classification and repair logic by constructing a business-driven dynamic data object context dictionary.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for cleaning edge data of industrial equipment integrated into MES, comprising: The edge computing unit obtains the production configuration database of MES in real time, acquires the inventory unit attribute parameter information corresponding to the current production order, and constructs a dynamic data object context dictionary based on the inventory unit attribute parameter information; the edge computing unit uses semantic mapping operators to structurally encapsulate the original sensor data objects of industrial field equipment, and transforms them into a standardized data stream containing equipment identifiers, timestamps and manufacturing status codes; A time-series sliding window is established in the edge computing unit. The statistical moment features and trend indicators of the standardized data stream within the time-series sliding window are extracted. Combined with the logical constraint thresholds in the context dictionary of the dynamic data object, a preset classification model is used to identify the current noise type. Based on the identification result of the noise type, the corresponding repair algorithm is dynamically matched from the preset cleaning strategy library to reconstruct the standardized data stream and generate high-fidelity data with business meaning. The high-fidelity data is mapped to the overall equipment efficiency calculation logic of MES, and the judgment weight parameters in the context-aware noise classification step are closed-loop iterated based on the feedback deviation of the overall equipment efficiency index.

[0008] Preferably, constructing the dynamic data object context dictionary includes: retrieving and extracting the process baseline value, physical quantity range, and production cycle constraint parameters corresponding to the current inventory unit from the production configuration database; establishing a mapping function between the inventory unit attribute parameters and the edge-side cleaning operator, converting the process baseline value into a dynamic statistical threshold for outlier determination, and converting the production cycle constraint parameters into a time window logic for identifying micro-downtime phenomena; associating the judgment criteria output by the mapping function with the corresponding device identifier, and writing them into the memory buffer of the edge processing unit in a key-value pair or structured object storage format to generate the dynamic data object context dictionary.

[0009] Preferably, the conversion to the standardized data stream includes: converting the raw sensor data objects acquired by the edge computing unit from binary analog quantities to floating-point digital physical quantities; calling the semantic mapping operator to assign a globally unique device identifier to each group of digital physical quantities according to the dynamic data object context dictionary, and injecting a high-precision timestamp using the system clock of the edge computing unit; extracting the I / O control status of industrial equipment in real time, mapping the current production operation mode to the corresponding manufacturing status code according to a preset logical mapping table, and appending the manufacturing status code as a dimension field to the digital physical quantity; combining the remapped device identifier, timestamp, manufacturing status code, and digital physical quantity into a preset structured object, and generating a standardized data stream with manufacturing context definition through memory serialization processing.

[0010] Preferably, the extraction of the statistical moment features and trend indicators includes: Within the time-series sliding window, digital physical quantities are extracted from the standardized data stream according to a preset step size, constructing a data sample sequence with a preset number of sampling points. Statistical operation functions are called to process the data sample sequence, extracting the arithmetic mean to characterize the central tendency of the data distribution, and statistical indicators to characterize the dispersion and skewness of the data distribution, as statistical moment features. A linear regression operator is used to analyze the evolution characteristics of the data sample sequence over time, extracting the slope index to characterize the drift velocity of physical quantities and the residual fluctuation index to characterize the fluctuation intensity, as trend indicators. The extracted statistical moment features and trend indicators are normalized and then concatenated to generate an input feature vector describing the current data window state characteristics.

[0011] Preferably, identifying the current noise type using a preset classification model includes: The generated input feature vector is input into the classification model and compared with the deviation of the normal production benchmark features pre-stored in the context dictionary of the dynamic data object. If the dispersion index in the input feature vector exceeds the statistical standard deviation by a preset multiple and the slope index is in an instantaneous jump state, the current noise type is determined to be a random impulse outlier. If the slope index in the input feature vector remains in the same direction within multiple preset continuous time windows, and the residual fluctuation index is within a preset stable range, then the current noise type is determined to be sensor trend drift; if the dispersion index in the input feature vector is less than a preset zero-point deviation threshold, and the digital physical quantity in the standardized data stream remains unchanged within a preset time step, then the current noise type is determined to be constant value deadlock.

[0012] Preferably, generating the high-fidelity data includes: retrieving the corresponding repair operator index from the cleaning strategy library based on the noise type; if the label is a random pulse outlier, invoking the linear interpolation operator; if the label is sensor trend drift, invoking the recursive estimation repair operator; if the label is normal data, executing the pass-through instruction; for data objects identified as random pulse outliers, performing linear interpolation operations using adjacent valid sampling points within the time-series sliding window to replace abnormal values; for data objects identified as sensor trend drift, using the recursive estimation repair operator to perform weighted fusion of the previous time-series state estimate and the current observation value to restore the physical true value; re-aligning and encapsulating the reconstructed value with the device identifier, timestamp, and manufacturing status code in the context dictionary of the dynamic data object; and combining the encapsulated structured objects into a high-fidelity data stream and storing it in the output buffer on the edge side.

[0013] Preferably, the judgment weight parameters in the closed-loop iterative context-aware noise classification step include: establishing a digital mapping relationship between the overall equipment efficiency index in the MES and the cleaning logic, and associating equipment downtime, performance deceleration, and quality defect indicators with the corresponding noise judgment weight factors; obtaining the actual output statistics fed back by the MES and the estimated value calculated based on the cleaning data, and calculating the response error and false alarm rate deviation between the two; if the response error and / or false alarm rate deviation exceeds the preset allowable range, calling the optimization algorithm to perform incremental compensation or penalty deduction on the judgment weight parameters in the classification model; for the working condition of frequent micro-shutdowns, automatically increasing the time threshold weight in the outlier noise judgment sub-step; and pushing the iterated judgment weight parameters to the execution environment on the edge side in real time, so as to realize the dynamic performance optimization of the cleaning logic without interrupting the data flow processing.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This application constructs a dynamic data object context dictionary, enabling the cleaning operator to perceive normal changes in physical quantities caused by the switching between production conditions and inventory units. This solves the problem of misjudgment caused by the lack of business context in traditional algorithms and ensures the semantic consistency of data cleaning under complex conditions.

[0015] 2. This application utilizes statistical moment features and trend indicators to construct multi-dimensional feature vectors, which, combined with a lightweight classification and recognition model, can accurately identify hidden noise such as random pulses, trend drift, and constant value hangs. Combined with edge-side repair operators, while ensuring high-fidelity data reconstruction, the processing latency is controlled at the millisecond level, meeting the needs of industrial real-time control.

[0016] 3. This application achieves continuous self-optimization of data processing logic by dynamically iterating the edge-side discrimination weight parameters through the deviation feedback of the overall equipment efficiency index of the MES terminal; it significantly reduces communication bandwidth costs and significantly improves the reliability of the MES decision support system. Attached Figure Description

[0017] Figure 1 This is a flowchart of an industrial equipment edge data cleaning method integrated into MES proposed in this invention. Figure 2 This is a flowchart of the method for identifying the current noise type proposed in this invention; Figure 3 This is a schematic diagram of the structure of an edge data cleaning method for industrial equipment integrated into MES proposed in this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0019] Please see Figures 1 to 3 This invention provides a method for cleaning edge data of industrial equipment integrated into MES, the technical solution of which is as follows: A method for cleaning edge data of industrial equipment integrated into MES, such as Figures 1-2 As shown, it includes: The edge computing unit obtains the production configuration database of MES in real time, acquires the inventory unit attribute parameter information corresponding to the current production order, and constructs a dynamic data object context dictionary based on the inventory unit attribute parameter information; the edge computing unit uses semantic mapping operators to structurally encapsulate the original sensor data objects of industrial field equipment, and transforms them into a standardized data stream containing equipment identifiers, timestamps and manufacturing status codes; A time-series sliding window is established in the edge computing unit. The statistical moment features and trend indicators of the standardized data stream within the time-series sliding window are extracted. Combined with the logical constraint thresholds in the context dictionary of the dynamic data object, a preset classification model is used to identify the current noise type. Based on the identification result of the noise type, the corresponding repair algorithm is dynamically matched from the preset cleaning strategy library to reconstruct the standardized data stream and generate high-fidelity data with business meaning. The high-fidelity data is mapped to the overall equipment efficiency calculation logic of MES, and the judgment weight parameters in the context-aware noise classification step are closed-loop iterated based on the feedback deviation of the overall equipment efficiency index.

[0020] Furthermore, constructing the dynamic data object context dictionary includes: retrieving and extracting the process baseline value, physical quantity range, and production cycle constraint parameters corresponding to the current inventory unit from the production configuration database; establishing a mapping function between the inventory unit attribute parameters and the edge-side cleaning operator, converting the process baseline value into a dynamic statistical threshold for outlier determination, and converting the production cycle constraint parameters into a time window logic for identifying micro-downtime phenomena; associating the determination criteria output by the mapping function with the corresponding device identifier, and writing them into the memory buffer of the edge processing unit in a key-value pair or structured object storage format to generate the dynamic data object context dictionary.

[0021] Specifically, the process of establishing the mapping function includes: defining the production cycle time constraint parameter as the target cycle time. And based on the sampling frequency of the edge computing unit The corresponding time window length threshold is calculated. ,in The preset volatility factor; when the data step size of consecutive zero volatility exceeds At that time, the logic item is activated to determine that the device is in a micro-shutdown state.

[0022] When constructing the dynamic statistical threshold, the mapping function extracts the process baseline value. And combined with the standard deviation within the historical window Generate dynamic judgment interval This serves as the logical boundary for outlier detection. The structured object contains a globally unique index, used to instantaneously address and call the corresponding memory buffer parameters based on the device identifier when the edge computing unit receives the standardized data stream, thereby achieving real-time synchronization between the cleaning logic and the current production task.

[0023] Specifically, the dynamic data object context dictionary is stored in the edge processing unit in the form of a hash table in memory, using the device identifier as the key for linear indexing. The structured object corresponding to each key contains necessary fields such as process baseline value (floating-point), upper limit of range (floating-point), lower limit of range (floating-point), production cycle threshold (integer), and update timestamp (long integer).

[0024] To ensure data consistency under millisecond-level query requirements, the dictionary employs a read-write lock queue mechanism: when the edge processing unit receives a production configuration update pushed by the MES, the system constructs a new version of the dictionary object in the background buffer and updates the current memory buffer through an atomic pointer switching operation. In scenarios with concurrent access from multiple devices, preset concurrency control operators ensure that the dictionary data is not corrupted during the addressing process of the cleaning operator, thereby guaranteeing read stability and sub-millisecond query response under high-frequency data stream impacts, meeting the real-time cleaning needs of industrial sites.

[0025] This invention achieves dynamic decoupling between cleaning thresholds and production tasks by constructing a dynamic data object context dictionary. An instantaneous addressing mechanism implemented using a globally unique index ensures millisecond-level synchronization of cleaning logic during task switching, eliminating the risk of false alarms from static thresholds. Simultaneously, production cycle time is transformed into quantified time window logic, accurately distinguishing between production intermittents and minor downtimes, avoiding metric distortion, and significantly improving the statistical accuracy of the MES. Furthermore, the adoption of a key-value pair storage format optimizes edge-side addressing performance, reduces computational overhead, and ensures real-time processing under high-frequency sampling.

[0026] Further, the transformation into the standardized data stream includes: converting the raw sensor data objects acquired by the edge computing unit from binary analog quantities to floating-point digital physical quantities; invoking the semantic mapping operator to assign a globally unique device identifier to each group of digital physical quantities according to the dynamic data object context dictionary, and injecting a high-precision timestamp using the system clock of the edge computing unit; extracting the I / O control status of industrial equipment in real time, mapping the current production operation mode to the corresponding manufacturing status code according to a preset logical mapping table, and appending the manufacturing status code as a dimension field to the digital physical quantity; combining the remapped device identifier, timestamp, manufacturing status code, and digital physical quantity into a preset structured object, and generating a standardized data stream with manufacturing context definition through memory serialization processing.

[0027] Specifically, the semantic mapping operator takes a sensor data object containing raw binary analog quantities as input and outputs a structured JSON object with manufacturing context definition. The "semantic" feature of the semantic mapping operator is reflected in the fact that it not only performs simple physical encapsulation, but also automatically converts the original physical attributes such as I / O channel numbers and register addresses into semantic tags with business meaning, such as field names like "spindle speed" and "bonding head pressure," by retrieving the context dictionary of the dynamic data object. Furthermore, it performs linear proportional conversion based on the range of physical quantities in the dictionary, mapping the binary analog quantities into floating-point digital physical quantities with physical units.

[0028] Specifically, the preset logic mapping table is stored in the static configuration area of ​​the edge computing unit, which defines the logical correspondence between multiple binary I / O level combinations and specific manufacturing status codes; for example, when it is detected that the I / O bit representing the spindle speed of the equipment is high and the I / O bit representing the emergency stop signal is low, the logic mapping table maps it to a manufacturing status code representing "normal production".

[0029] During the memory serialization process, the edge computing unit uses a preset structured byte alignment protocol to arrange the device identifier, timestamp, manufacturing status code, and floating-point digital physical quantities in a continuous memory space according to a fixed byte offset. By adding a message header containing the total data length and checksum to the memory block header, the scattered dimension fields are converted into a binary bit stream that can be directly addressed and read by the algorithm model, thus generating the standardized data stream.

[0030] This invention uses semantic mapping operators and logical mapping tables to associate low-level binary signals with high-level manufacturing status codes in real time, achieving accurate conversion of raw sensor data into standardized data streams with business semantics. By utilizing a structured byte alignment protocol and message header verification mechanism, it ensures continuous storage and fast addressing of heterogeneous dimensional fields in memory, significantly improving the algorithm model's performance in reading time-series data and the integrity of its transmission. This processing flow not only provides high-precision structured input for edge-side classification and recognition but also ensures the traceability of data throughout the entire manufacturing process through unified device identification and timestamp injection.

[0031] Furthermore, the extraction of the statistical moment features and trend indicators includes: Within the time-series sliding window, digital physical quantities are extracted from the standardized data stream according to a preset step size, constructing a data sample sequence with a preset number of sampling points. Statistical operation functions are called to process the data sample sequence, extracting the arithmetic mean to characterize the central tendency of the data distribution, and statistical indicators to characterize the dispersion and skewness of the data distribution, as statistical moment features. A linear regression operator is used to analyze the evolution characteristics of the data sample sequence over time, extracting the slope index to characterize the drift velocity of physical quantities and the residual fluctuation index to characterize the fluctuation intensity, as trend indicators. The extracted statistical moment features and trend indicators are normalized and then concatenated to generate an input feature vector describing the current data window state characteristics.

[0032] Specifically, the preset step size It is based on the real-time computing load of the edge computing unit and the sampling frequency of the raw sensor data. It is dynamically determined, and its value range is set as follows: ,in The total duration of the time-series sliding window. The preset number of sampling points is used. In practical applications, the step size is usually a fixed integer multiple of the sampling period between 10ms and 50ms to ensure that when constructing a data sample sequence with the preset number of sampling points, the complete production cycle fluctuation characteristics can be covered without aliasing.

[0033] When the production cycle constraint parameters in the context dictionary of the dynamic data object change, the edge computing unit will automatically adjust the step size. This ensures that the number of sample points extracted within a production cycle remains constant, thereby guaranteeing the comparability of the subsequently extracted arithmetic mean and slope indicators under different working conditions.

[0034] Specifically, in the process of extracting statistical moment features, the arithmetic mean, standard deviation, and skewness coefficient of the data sample sequence are calculated by calling first- to third-order cumulant operators, respectively; wherein, the skewness coefficient is used to quantify the degree of asymmetry of the distribution of the digital physical quantity relative to the mean, so as to identify asymmetric impulse interference of the sensor signal.

[0035] When extracting the trend indicators, the linear regression operator uses the least squares method to fit the data sample sequence, calculates the slope of the fitted line as the slope indicator, and calculates the root mean square error of the sample points relative to the fitted line as the residual fluctuation indicator. The normalization process uses the minimum-maximum normalization operator to scale the statistical moment features and trend indicators of different dimensions to a unified numerical range of [0, 1]. The cascading process uses a vector concatenation operator to arrange the normalized indicators of each dimension in a preset order, generating a fixed-length input feature vector for logical discrimination in subsequent classification models.

[0036] This invention achieves in-depth quantification of the distribution characteristics and evolution trends of industrial data through the collaborative extraction of multi-order cumulant operators and least squares regression operators. The introduction of skewness coefficients and residual fluctuation indices significantly enhances the system's ability to characterize asymmetric interference and hidden drift noise. Combining max-min normalization and vector concatenation not only eliminates the interference of different physical dimensions on model judgments but also improves the classification model's discriminative power and noise resistance in edge computing environments by constructing high-dimensional feature vectors, laying a solid logical foundation for accurately identifying complex mixed noise.

[0037] Furthermore, such as Figure 3 As shown, the current noise type is identified using a preset classification model, including: The generated input feature vector is input into the classification model and compared with the deviation of the normal production benchmark features pre-stored in the context dictionary of the dynamic data object. If the dispersion index in the input feature vector exceeds the statistical standard deviation by a preset multiple and the slope index is in an instantaneous jump state, the current noise type is determined to be a random impulse outlier. If the slope index in the input feature vector remains in the same direction within multiple preset continuous time windows, and the residual fluctuation index is within a preset stable range, then the current noise type is determined to be sensor trend drift; if the dispersion index in the input feature vector is less than a preset zero-point deviation threshold, and the digital physical quantity in the standardized data stream remains unchanged within a preset time step, then the current noise type is determined to be constant value deadlock.

[0038] The preset classification model employs a lightweight fully connected neural network, deployed within the inference engine of an edge computing unit. In terms of model architecture, the input layer has a fixed 5-dimensional dimension, corresponding to the arithmetic mean, standard deviation, skewness coefficient, slope index, and residual fluctuation index in the input feature vector. The hidden layer uses a single-layer structure with 16 neurons, uniformly employing the ReLU activation function for non-linear feature mapping. The output layer uses the Softmax function, mapping the output features to probability scores corresponding to four categories of labels: random impulse outliers, sensor trend drift, constant value deadlock, and normal data.

[0039] Regarding the model training mechanism, the system utilizes historical operational data from industrial sites to construct a labeled training set. This dataset is generated by synchronously acquiring raw sensor data from the production line under known interference conditions, and manually labeling the corresponding noise types based on physical experimental phenomena. The stochastic gradient descent algorithm is used to train the model weights offline until the model's classification accuracy on the validation set reaches the preset target.

[0040] Specifically, the deviation comparison is achieved by calling the Euclidean distance operator to calculate the spatial distance between the input feature vector and the reference feature vector as the deviation metric. When determining random pulse outliers, the statistical standard deviation of the preset multiple is set to 3 times the standard deviation; the instantaneous jump state refers to the absolute value of the difference between two adjacent sampling points exceeding 5 times the average amplitude of the standardized data stream over the past preset period.

[0041] To ensure real-time inference at the edge, the classification model underwent weight quantization before deployment, compressing floating-point parameters into 8-bit integer parameters. The model's recognition logic follows the principle of "consistency between inference judgment and rule verification": when the probability score of a certain noise category in the output layer is the highest, and the feature of that dimension simultaneously meets statistical quantification standards (such as dispersion exceeding 3 times the standard deviation), the model outputs a determined noise type label.

[0042] When determining sensor trend drift, the same-direction offset is achieved by calculating the sign consistency of the slope index within M consecutive windows, where M is an integer not less than 5; the stable interval is defined as the rate of change of the residual fluctuation index within a preset percentage range. When determining constant value deadlock, the zero-point deviation threshold is set according to 1% of the sensor's physical resolution; by setting a judgment score function in the classification model, the judgment results of each feature index are weighted and summed, and when the score exceeds the preset classification threshold, the corresponding noise type label is output.

[0043] This invention achieves accurate classification and quantitative identification of typical noise patterns in industrial environments by logically coupling the Euclidean distance operator with multi-dimensional decision indicators. It effectively distinguishes complex anomalies such as instantaneous pulses, hidden drifts, and state hangs, solving the problem of traditional algorithms having a single identification dimension at the edge. Through the weighted output of the decision scoring function, it significantly reduces the false alarm rate caused by environmental interference, provides a highly reliable logical index for backend data reconstruction, and enhances the system's decision-making stability under heterogeneous operating conditions.

[0044] Further, generating the high-fidelity data includes: retrieving the corresponding repair operator index from the cleaning strategy library based on the noise type; if the label is a random impulse outlier, calling the linear interpolation operator; if the label is sensor trend drift, calling the recursive estimation repair operator; if the label is a constant value stuck, calling the forward filling operator to replace the constant value during the stuck period with the value of the last valid sampling point; if the label is normal data, executing the pass-through instruction; for data objects identified as random impulse outliers, performing linear interpolation operations using adjacent valid sampling points within the time-series sliding window to replace the abnormal values; for data objects identified as sensor trend drift, using the recursive estimation repair operator to perform weighted fusion of the state estimate value of the previous moment and the current observation value to restore the physical true value; re-aligning and encapsulating the reconstructed value with the device identifier, timestamp, and manufacturing status code in the context dictionary of the dynamic data object; combining the encapsulated structured objects into a high-fidelity data stream and storing it in the output buffer on the edge side.

[0045] Specifically, the linear interpolation operation obtains two valid sampling points before and after the abnormal value. and The slope is calculated and compensated according to the time step to generate reconstructed values. This ensures the continuity of the time-series trajectory. During the recursive estimation and repair process, the operator employs first-order lag filtering logic or Kalman gain update logic. By calculating the prediction residual at the current time step, it dynamically adjusts the weighting ratio between the previous time step's state estimate and the current observation. Wherein, the weighting ratio The system automatically calculates the system process noise covariance and observation noise covariance based on preset parameters to achieve a smooth restoration of sensor trend drift.

[0046] During the realignment and encapsulation process, the edge computing unit uses a structured alignment algorithm to merge the reconstructed values ​​with its original device identifier, timestamp, and manufacturing status code in memory according to a preset field offset. By calculating the cyclic redundancy check code of the encapsulated structured object and appending it to the end of the data, the integrity and parsing consistency of the high-fidelity data stream are ensured during storage in the output buffer and subsequent transmission to the MES.

[0047] The recursive estimation and repair operator is implemented using first-order hysteresis filtering logic. During weighted fusion, the system calculates the reconstructed value at the current moment using a preset smoothing coefficient. This reconstructed value is composed of the state estimate from the previous moment and the current sensor observation, proportionally allocated. This proportion is not a fixed value but is dynamically determined based on the range in the dynamic data object's context dictionary and the trend indicator in the current input feature vector: when the trend indicator shows an accelerated drift, the system automatically lowers the weight ratio of the current observation and increases the reference proportion of the previous moment's estimate, thereby effectively suppressing noise interference in the observation data. In actual industrial edge computing scenarios, the preset smoothing coefficient is typically set between 0.05 and 0.3.

[0048] This invention achieves accurate reconstruction and physical feature restoration of abnormal data through a classification matching repair operator and a dynamic weighted fusion mechanism. By utilizing linear interpolation and recursive estimation logic, the impact of impulse noise and trend drift on data quality is effectively eliminated, ensuring the continuity of time-series trajectories. Combined with a structured alignment algorithm and cyclic redundancy check, the reconstructed data ensures high transmission integrity and parsing consistency while preserving the manufacturing context, thereby outputting a high-fidelity data stream with business semantics to the MES, significantly improving the reliability of production decisions.

[0049] Furthermore, the judgment weight parameters in the closed-loop iterative context-aware noise classification step include: establishing a digital mapping relationship between the overall equipment efficiency index in the MES and the cleaning logic, and associating equipment downtime, performance deceleration, and quality defect indicators with the corresponding noise judgment weight factors; obtaining the actual output statistics fed back by the MES and the estimated value calculated based on the cleaning data, and calculating the response error and false alarm rate deviation between the two; if the response error and / or false alarm rate deviation exceeds the preset allowable range, calling the optimization algorithm to perform incremental compensation or penalty deduction on the judgment weight parameters in the classification model; for the working condition of frequent micro-shutdowns, automatically increasing the time threshold weight in the outlier noise judgment sub-step; and pushing the iterated judgment weight parameters to the edge execution environment in real time, so as to realize the dynamic performance optimization of the cleaning logic without interrupting the data flow processing.

[0050] Specifically, the digital mapping relationship is achieved through a preset weighted correlation matrix, where the equipment downtime index is correlated to the duration threshold factor for outlier determination, and the quality defect index is correlated to the variance threshold factor for drift determination. The response error is calculated by taking the actual output statistics of the MES. Compared with the estimated value The normalized residuals between them are used to achieve the false alarm rate deviation; the false alarm rate deviation is obtained by comparing the overlap between the abnormal events determined by the edge side and the real fault labels after manual review.

[0051] When performing incremental compensation or penalty deduction of the judgment weight parameters, an optimization algorithm based on gradient descent or incremental proportional integral operator is invoked, according to the formula. The weight parameters are iterated, where The preset learning rate or compensation step size, This is to account for changes in response error or false alarm rate deviation. For micro-shutdown scenarios, the system dynamically increases the time window width coefficient in outlier detection by identifying production interruption signals between 200ms and 2s, to prevent instantaneous jitter from being misjudged as equipment shutdown. The model hot update is achieved through a dual-buffer switching mechanism of the edge computing unit, ensuring that new weight parameters take effect instantly via pointer switching after being loaded in the background, guaranteeing the continuity of data stream processing.

[0052] The optimization algorithm employs incremental proportional-integral (PI) adjustment logic. This algorithm takes the response error from the MES feedback as input and converts the error into a correction value for the judgment weight based on preset proportional and integral coefficients. In terms of weight adjustment strategy, the system prioritizes incremental compensation for the single-dimensional weight with the highest contribution to the response error. If multiple indicators simultaneously deviate from the preset range, punitive deductions are performed sequentially with fixed step sizes according to the priority order defined in the weight correlation matrix.

[0053] Regarding the termination and convergence conditions of the iteration, the optimization algorithm has a built-in dual trigger threshold: the first is the response error threshold, when the statistical error of three consecutive production cycles is less than a preset percentage, the system determines that the model has reached the convergence state and stops updating the weights; the second is the maximum number of iterations limit, which is preset to fifty times, to prevent algorithm oscillation under extreme fluctuation conditions.

[0054] This invention achieves adaptive evolution and accuracy optimization in the data governance process by establishing a closed-loop feedback mechanism between business indicators and data cleaning logic. Utilizing a weighted correlation matrix and an incremental iterative algorithm, the discrimination criteria can be dynamically adjusted based on production performance deviations, effectively solving technical challenges such as misjudgments of minor downtime under complex operating conditions. Combined with a dual-buffer hot update mechanism, it ensures seamless switching and continuous operation of the system during parameter tuning, significantly enhancing the robustness of edge computing units and guaranteeing the long-term stable operation of the manufacturing system.

[0055] This invention constructs an industrial data governance architecture with manufacturing context awareness through deep decoupling and closed-loop collaboration between edge computing and MES business logic. By using a dynamic data object context dictionary and semantic mapping operators, it eliminates the risk of misjudgment caused by the disconnect between traditional cleaning algorithms and production conditions, ensuring the business authenticity of the data stream. Utilizing pattern recognition logic based on multidimensional statistical moment features and trend indicators, it achieves accurate classification and high-fidelity reconstruction of complex mixed noise, significantly improving the quality of the underlying data. The most crucial benefit lies in the fact that by driving the self-iteration of judgment weight parameters through feedback deviations in overall equipment efficiency indicators, the cleaning strategy can dynamically evolve with equipment status and process requirements, significantly reducing network bandwidth pressure while providing highly reliable decision-making basis for MES. Example

[0056] In practical industrial applications, taking a semiconductor precision packaging production line as an example, the specific implementation process of the industrial equipment edge data cleaning method integrated into the manufacturing execution system of this invention is as follows: The edge computing unit accesses the manufacturing execution system's production configuration database in real time, retrieving inventory unit attribute parameters such as packaging baseline pressure, bonding head temperature range, and production cycle time corresponding to the current wafer bonding production order. Based on these inventory unit attribute parameters, the edge computing unit constructs a dynamic data object context dictionary. Specifically, it transforms the process baseline pressure into a dynamic statistical threshold for identifying pressure outliers and the production cycle time into a time window logic for identifying micro-stops in the bonding process, storing these parameters in a memory buffer as key-value pairs. Simultaneously, the edge computing unit acquires the binary pressure signal generated by the bonding machine's raw sensors, converts it into a floating-point digital physical quantity, and injects the bonding machine's device identifier and high-precision timestamp using a semantic mapping operator. The system further extracts the bonding machine's control state, mapping the current operating mode to a specific manufacturing status code, and finally encapsulates it into a standardized data stream containing manufacturing context definitions through memory serialization.

[0057] Within a time-series sliding window established in the edge computing unit, the system extracts the pressure-standardized data stream by step size and constructs a sampling sequence. The mean, standard deviation, and skewness coefficient of this sequence are calculated using statistical functions as statistical moment features. Simultaneously, a linear regression operator is used to calculate the slope index and residual fluctuation index of pressure evolution as trend indicators. These normalized indicators are concatenated to generate an input feature vector, and a pre-defined classification model is used to identify noise types. If the pressure index exceeds three times the standard deviation and exhibits instantaneous jumps, it is identified as a random impulse outlier; if the slope index maintains a consistent directional shift across multiple consecutive windows and the residual fluctuation is stable, it is identified as sensor trend drift.

[0058] Based on the identified noise type, the system matches a repair algorithm from the cleaning strategy library. For pressure pulse interference, a linear interpolation operator is invoked to replace outliers with adjacent valid sampling points. For pressure sensor drift, a recursive estimation repair operator is used to weightedly fuse the estimated and observed values ​​to restore the true bonding pressure value. The reconstructed pressure value is then re-aligned and encapsulated with the original equipment identifier, timestamp, and manufacturing status code to generate a high-fidelity data stream with business meaning, which is then stored in the output buffer.

[0059] Finally, the cleaned high-fidelity pressure data is mapped into the overall equipment efficiency calculation logic of the Manufacturing Execution System (MES). The MES feeds back to the edge computing unit based on the response error between the actual bonding output and the estimated value based on the cleaned data, as well as the false alarm rate deviation from manual review. The edge computing unit calls an optimization algorithm to perform incremental compensation iterations on the noise determination weight parameters and utilizes a dual-buffer switching mechanism to achieve hot updates of the determination weight parameters, thereby continuously optimizing the cleaning accuracy without interrupting the bonding pressure data processing.

[0060] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for cleaning edge data of industrial equipment integrated into MES, characterized in that, include: The edge computing unit obtains the production configuration database of MES in real time, acquires the inventory unit attribute parameter information corresponding to the current production order, and constructs a dynamic data object context dictionary based on the inventory unit attribute parameter information; the edge computing unit uses semantic mapping operators to structurally encapsulate the original sensor data objects of industrial field equipment, and transforms them into a standardized data stream containing equipment identifiers, timestamps and manufacturing status codes; A time-series sliding window is established in the edge computing unit. The statistical moment features and trend indicators of the standardized data stream within the time-series sliding window are extracted. Combined with the logical constraint threshold in the context dictionary of the dynamic data object, the current noise type is identified using a preset classification model. Based on the noise type identification result, the corresponding repair algorithm is dynamically matched from the preset cleaning strategy library to reconstruct the standardized data stream and generate high-fidelity data with business meaning; The high-fidelity data is mapped to the overall equipment efficiency calculation logic of MES, and the judgment weight parameters in the context-aware noise classification step are closed-loop iterated based on the feedback deviation of the overall equipment efficiency index.

2. The method for cleaning edge data of industrial equipment integrated into MES according to claim 1, characterized in that, Constructing the dynamic data object context dictionary includes: retrieving and extracting the process baseline value, physical quantity range, and production cycle constraint parameters corresponding to the current inventory unit from the production configuration database; establishing a mapping function between the inventory unit attribute parameters and the edge-side cleaning operator, converting the process baseline value into a dynamic statistical threshold for outlier determination, and converting the production cycle constraint parameters into a time window logic for identifying micro-downtime phenomena; associating the determination criteria output by the mapping function with the corresponding device identifier, and writing them into the memory buffer of the edge processing unit in a key-value pair and / or structured object storage format to generate the dynamic data object context dictionary.

3. The method for cleaning edge data of industrial equipment integrated into MES according to claim 1, characterized in that, The transformation into the standardized data stream includes: converting the raw sensor data objects acquired by the edge computing unit from binary analog quantities to floating-point digital physical quantities; calling the semantic mapping operator to assign a globally unique device identifier to each group of digital physical quantities according to the dynamic data object context dictionary, and injecting a high-precision timestamp using the system clock of the edge computing unit; extracting the I / O control status of industrial equipment in real time, mapping the current production operation mode to the corresponding manufacturing status code according to the preset logical mapping table, and appending the manufacturing status code as a dimension field to the digital physical quantity; combining the remapped device identifier, timestamp, manufacturing status code, and digital physical quantity into a preset structured object, and generating a standardized data stream with manufacturing context definition through memory serialization processing.

4. The method for cleaning edge data of industrial equipment integrated into MES according to claim 1, characterized in that, Extracting the statistical moment features and trend indicators includes: Within the time-series sliding window, digital physical quantities are extracted from the standardized data stream according to a preset step size, constructing a data sample sequence with a preset number of sampling points. Statistical operation functions are called to process the data sample sequence, extracting the arithmetic mean to characterize the central tendency of the data distribution, and statistical indicators to characterize the dispersion and skewness of the data distribution, as statistical moment features. A linear regression operator is used to analyze the evolution characteristics of the data sample sequence over time, extracting the slope index to characterize the drift velocity of physical quantities and the residual fluctuation index to characterize the fluctuation intensity, as trend indicators. The extracted statistical moment features and trend indicators are normalized and then concatenated to generate an input feature vector describing the current data window state characteristics.

5. The method for cleaning edge data of industrial equipment integrated into MES according to claim 1, characterized in that, Identifying the current noise type using a pre-defined classification model includes: The generated input feature vector is input into the classification model and compared with the deviation of the normal production benchmark features pre-stored in the context dictionary of the dynamic data object. If the dispersion index in the input feature vector exceeds the statistical standard deviation by a preset multiple and the slope index is in an instantaneous jump state, the current noise type is determined to be a random impulse outlier. If the slope index in the input feature vector remains in the same direction within multiple preset continuous time windows, and the residual fluctuation index is within a preset stable range, then the current noise type is determined to be sensor trend drift; if the dispersion index in the input feature vector is less than a preset zero-point deviation threshold, and the digital physical quantity in the standardized data stream remains unchanged within a preset time step, then the current noise type is determined to be constant value deadlock.

6. The method for cleaning edge data of industrial equipment integrated into MES according to claim 1, characterized in that, Generating the high-fidelity data includes: retrieving the corresponding repair operator index from the cleaning strategy library based on the noise type; if the label is a random pulse outlier, calling the linear interpolation operator; if the label is sensor trend drift, calling the recursive estimation repair operator; if the label is normal data, executing the pass-through command; for data objects identified as random pulse outliers, performing linear interpolation operations using adjacent valid sampling points within the time-series sliding window to replace abnormal values; for data objects identified as sensor trend drift, using the recursive estimation repair operator to perform weighted fusion of the previous time-series state estimate and the current observation value to restore the physical true value; re-aligning and encapsulating the reconstructed value with the device identifier, timestamp, and manufacturing status code in the context dictionary of the dynamic data object; combining the encapsulated structured objects into a high-fidelity data stream and storing it in the output buffer on the edge side.

7. The method for cleaning edge data of industrial equipment integrated into MES according to claim 1, characterized in that, The judgment weight parameters in the closed-loop iterative context-aware noise classification step include: establishing a digital mapping relationship between the overall equipment efficiency index in the MES and the cleaning logic, and associating equipment downtime, performance deceleration, and quality defect indicators with the corresponding noise judgment weight factors; obtaining the actual output statistics fed back by the MES and the estimated value calculated based on the cleaning data, and calculating the response error and false alarm rate deviation between the two; if the response error and / or false alarm rate deviation exceeds the preset allowable range, calling the optimization algorithm to perform incremental compensation and / or penalty deduction on the judgment weight parameters in the classification model; for the working condition of frequent micro-shutdowns, automatically increasing the time threshold weight in the outlier noise judgment sub-step; and pushing the iterated judgment weight parameters to the edge execution environment in real time, so as to realize the dynamic performance optimization of the cleaning logic without interrupting the data flow processing.

Images