Self-diagnosis and self-calibration method for long-term drift of powder quantitative packaging precision

Abstract

The application provides a powder quantitative packaging precision long-term drift self-diagnosis and self-calibration method, comprising: aiming at the influence of multiple drift sources such as screw wear, environmental humidity change, pneumatic response lag and the like on process precision, a static parameter drift causal skeleton diagram combining mechanical structure topology and process constraints is constructed, and through collecting multi-source heterogeneous time sequence data, normalization and feature extraction are carried out to form a quantifiable observation evidence set; a causal reasoning diagram with mechanism weight and uncertainty parameter is used, an improved A* algorithm is used to dynamically optimize the tracing path, an interpretable correction decision tree is generated, and a structured correction instruction package is output; after human-computer interface interaction confirmation, automatic correction and closed-loop calibration of target process parameters are realized, and the application improves the accuracy of equipment fault diagnosis and the timeliness of correction, and realizes effective tracing and self-adaptive regulation and control of parameter drift.

Description

Technical Field

[0001] This invention relates to the field of automated powder quantitative packaging and equipment self-calibration technology, and in particular to a self-diagnosis and self-calibration method for long-term drift in powder quantitative packaging accuracy. Background Technology

[0002] The widespread application of automated powder quantitative packaging equipment in industrial production has promoted continuous improvement in packaging efficiency and weighing accuracy. Current mainstream powder packaging equipment self-calibration and parameter correction technologies mainly employ periodic or trigger-based monitoring schemes. These rely on single or simple integrated signals such as weighing residuals, batch error statistics, and environmental parameters, combined with threshold logic, empirical rules, and some black-box machine learning models (such as traditional discriminant classification algorithms) to determine several possibilities of parameter drift. In industrial control systems, the central control HMI interface typically only provides simple prompts for calibration actions and results in unstructured ways such as alarms and parameter reports, lacking systematic explanations of the diagnostic process and visual feedback of causes. In recent years, research interest in the automatic calibration of packaging equipment parameters has increased, and industry development trends are reflected in two aspects: On the one hand, attempts are being made to incorporate more sensor data (such as motor current, ambient temperature and humidity, and material feeding images) into data fusion diagnostics to improve self-calibration accuracy. However, data fusion often relies on weighted averaging, simple discrimination, or static logic combination. On the other hand, some high-end equipment has begun to introduce machine learning discrimination models (such as decision trees and shallow neural networks) for parameter anomaly detection. However, most of these models are black-box predictions, and their outputs are only correction suggestions. Humans cannot trace the rationale behind their decisions, making it difficult to trace the physical and data basis of each step in the parameter correction process. Furthermore, existing technologies cannot meet the core requirements of modern intelligent manufacturing for "explainable, traceable, interactive, and verifiable self-calibration processes." For maintenance personnel, identifying the root causes of parameter drift, verifying the rationality of correction suggestions, and quickly locating potential problems during long-term operation is extremely difficult, leading to high misdiagnosis rates, low system trust, and increased costs for subsequent maintenance and optimization. In complex equipment scenarios (such as multi-source drift collaboration, sudden event interference, and structural changes), the lack of a parameter correction visualization mechanism that can integrate "data, causal mechanisms, and human-machine interaction" easily creates bottlenecks such as difficulty in auditing the calibration process, difficulty in accumulating experience, and difficulty in iterative optimization. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a self-diagnosis and self-calibration method for long-term drift of quantitative packaging accuracy of powder.

[0004] The technical solution of this invention is achieved as follows: a self-diagnosis and self-calibration method for long-term drift in the quantitative packaging accuracy of powders, comprising: S1: Based on the mechanical structure topology and process physical constraints of powder quantitative packaging equipment, a static parameter drift causal skeleton diagram is constructed, which includes 12 typical drift sources such as screw wear, environmental humidity change, and aerodynamic response hysteresis. The causal skeleton diagram consists of drift source nodes, observable parameter nodes, and directional causal edges, serving as the basic structure for subsequent dynamic source tracing reasoning. S2: Collect four types of heterogeneous time-series data from the weighing module, drive motor encoder, environmental sensor and vision-assisted positioning module at each calibration trigger cycle, and perform sliding window segmentation and zero-mean normalization on the data of each channel to eliminate dimensional differences and generate a standardized time-series segment sequence. S3: For each type of standardized time series segment, extract three statistical features: kurtosis, skewness, and autocorrelation decay time constant to form a multi-dimensional feature vector; map each feature to the corresponding observable parameter node in the static parameter drift causal skeleton diagram to generate a multi-source observation evidence set with quantized intensity scalars; S4: Based on the strength scalar of each piece of evidence in the multi-source observation evidence set and the mechanism weight and physical uncertainty interval of the corresponding causal edge, update the confidence of each path in combination with the historical evidence decay rule, and trace the high-confidence tracing path that satisfies the optimal condition of the scoring function in reverse on the enhanced causal graph based on the improved A* search algorithm, and generate an interpretable corrected decision tree containing at least two causal nodes. S5: Transform the root node error, intermediate causal links and end parameter compensation suggestions in the interpretable correction decision tree into a structured correction instruction package with weighted labels and original data references. The compensation suggestions explicitly point to controllable process parameters such as the filling volume compensation coefficient or the material feeding delay threshold. S6: Before executing the correction instructions, the structured correction instruction package is input into the lightweight WebGL rendering engine to generate an interactive SVG format dynamic traceability map, which is then projected onto the human-machine interface in real time, allowing operators to view the supporting evidence waveforms, feature calculation basis, and recommended action prompts for each node. S7: If the operator confirms the parameter correction scheme guided by the dynamic traceability map on the human-machine interface, a structured correction instruction package is sent to the controller to drive the actuator to automatically adjust the target process parameters and complete the closed-loop self-calibration action. S8: Record the dynamic traceability graph structure, user interaction behavior logs, and residual change data before and after calibration during this calibration process. Store the data in the local time-series knowledge graph database in the form of RDF triples. Update the prior weights of relevant edges in the causal skeleton graph dynamically based on human feedback results to achieve iterative optimization of the credibility of the inference model.

[0005] The self-diagnosis and self-calibration method for long-term drift of powder quantitative packaging accuracy provided by the present invention has the following beneficial effects: (1) This invention significantly improves the interpretability and decision transparency of the parameter correction process by constructing a parameter drift causal skeleton diagram based on physical mechanisms and process constraints. During system initialization, a mechanical topology association consisting of key components such as the feeding screw, weighing sensor, and pneumatic gate is established. Combining physical laws such as mass conservation, torque balance, and rheological hysteresis, 12 typical drift sources and their directional causal relationships with observable parameters are predefined, forming a static causal skeleton G0. This structure maps abnormal performance during equipment operation to a concrete fault propagation path, allowing each parameter shift to be traced back to a specific mechanical or environmental cause. For example, "decreased filling coefficient" clearly points to "screw wear," and "increased material feeding fluctuation" is attributed to "powder agglomeration." This knowledge-driven modeling approach fundamentally changes the previous correction logic that relied solely on data fitting, achieving a leap from "passive response" to "active attribution." (2) This invention introduces a dynamic injection of multi-source heterogeneous evidence and a graph-based decision generation mechanism, which effectively ensures the accuracy, real-time performance, and human-machine collaboration efficiency of parameter correction. In each calibration trigger cycle, the system synchronously collects four types of time-series data: weighing residual, motor encoder, ambient temperature and humidity, and visual positioning. It extracts high-order statistical features such as kurtosis, skewness, and autocorrelation decay time constant and converts them into evidence strength scalars on the causal edges. Combined with time decay rules, it updates historical confidence weights to ensure that recent observations have a dominant influence on the inference results. On this basis, an improved A* algorithm is used to search for the optimal causal path in reverse. Taking into account the mechanism weights, evidence strength, and physical uncertainty intervals set by experts, it outputs a high-confidence decision chain containing multiple causal nodes. For example, when the weighing residual is detected to be continuously positively biased and the motor current spectrum entropy value increases, the system can accurately infer that "increased feed screw clearance" is the core cause and recommend corresponding compensation strategies. More importantly, the entire reasoning process is presented in an interactive SVG-format traceability graph on the HMI interface. Operators can expand each node layer by layer to view the original waveforms, calculation formulas, historical cases, and handling suggestions, achieving a "what you see is what you get" visualized calibration process. All graph structures and user feedback are stored in a local time-series knowledge graph library as RDF triples, supporting semantic retrieval and long-term evolution analysis, and constructing a closed-loop, reliable iterative system of "perception-reasoning-execution-feedback". Attached Figure Description

[0006] Figure 1 This is a flowchart of the self-diagnosis and self-calibration method for long-term drift of powder quantitative packaging accuracy according to the present invention. Figure 2 This is a sub-flowchart of the self-diagnosis and self-calibration method for long-term drift of powder quantitative packaging accuracy according to the present invention; Figure 3 This is another sub-flowchart of the self-diagnosis and self-calibration method for long-term drift of powder quantitative packaging accuracy according to the present invention. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0008] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0009] like Figure 1 As shown, this invention provides a self-diagnosis and self-calibration method for long-term drift in the quantitative packaging accuracy of powders, specifically including: S1: Based on the mechanical structure topology and process physical constraints of powder quantitative packaging equipment, a static parameter drift causal skeleton diagram is constructed, which includes 12 typical drift sources such as screw wear, environmental humidity change, and aerodynamic response hysteresis. The causal skeleton diagram consists of drift source nodes, observable parameter nodes, and directional causal edges, serving as the basic structure for subsequent dynamic source tracing reasoning. S2: Collect four types of heterogeneous time-series data from the weighing module, drive motor encoder, environmental sensor and vision-assisted positioning module at each calibration trigger cycle, and perform sliding window segmentation and zero-mean normalization on the data of each channel to eliminate dimensional differences and generate a standardized time-series segment sequence. S3: For each type of standardized time series segment, extract three statistical features: kurtosis, skewness, and autocorrelation decay time constant to form a multi-dimensional feature vector; map each feature to the corresponding observable parameter node in the static parameter drift causal skeleton diagram to generate a multi-source observation evidence set with quantized intensity scalars; S4: Based on the strength scalar of each piece of evidence in the multi-source observation evidence set and the mechanism weight and physical uncertainty interval of the corresponding causal edge, update the confidence of each path in combination with the historical evidence decay rule, and trace the high-confidence tracing path that satisfies the optimal condition of the scoring function in reverse on the enhanced causal graph based on the improved A* search algorithm, and generate an interpretable corrected decision tree containing at least two causal nodes. S5: Transform the root node error, intermediate causal links and end parameter compensation suggestions in the interpretable correction decision tree into a structured correction instruction package with weighted labels and original data references. The compensation suggestions explicitly point to controllable process parameters such as the filling volume compensation coefficient or the material feeding delay threshold. S6: Before executing the correction instructions, the structured correction instruction package is input into the lightweight WebGL rendering engine to generate an interactive SVG format dynamic traceability map, which is then projected onto the human-machine interface in real time, allowing operators to view the supporting evidence waveforms, feature calculation basis, and recommended action prompts for each node. S7: If the operator confirms the parameter correction scheme guided by the dynamic traceability map on the human-machine interface, a structured correction instruction package is sent to the controller to drive the actuator to automatically adjust the target process parameters and complete the closed-loop self-calibration action. S8: Record the dynamic traceability graph structure, user interaction behavior logs, and residual change data before and after calibration during this calibration process. Store the data in the local time-series knowledge graph database in the form of RDF triples. Update the prior weights of relevant edges in the causal skeleton graph dynamically based on human feedback results to achieve iterative optimization of the credibility of the inference model.

[0010] Step S1: Based on the mechanical structure topology and process physical constraints of the powder quantitative packaging equipment, a static parameter drift causal skeleton diagram is constructed, including 12 typical drift sources such as screw wear, environmental humidity changes, and aerodynamic response hysteresis. This causal skeleton diagram consists of drift source nodes, observable parameter nodes, and directional causal edges, serving as the basic structure for subsequent dynamic source tracing and reasoning. Specifically, it includes: S1.1: Based on the mechanical structure topology of the powder quantitative packaging equipment, the physical connection sequence between the feeding screw, weighing sensor, pneumatic gate, and vibrating feeder is obtained. The mechanical coupling link matrix M is generated using the equipment assembly drawings and kinematic chain analysis methods, where each non-zero element... This indicates that the i-th component has a force or displacement transmission effect on the j-th component, which serves as the spatial basis for constructing the connectivity between nodes in the causal skeleton diagram; Based on the mechanical structure topology of the powder quantitative packaging equipment, an assembly drawing analysis method (parameters: CAD format drawing file, component identifier set) is used to realize the systematic analysis of the geometric position and connection sequence of the feeding screw, weighing sensor, pneumatic gate and vibrating feeder. Furthermore, by using the kinematic chain analysis method (parameters: constraint type set, revolute joint / prismatic joint definition), the force and displacement transmission relationship between each component is modeled, and a set of initial mechanical coupling link graph data structure containing nodes and connecting edges is obtained; Furthermore, a matrix encoding method (parameters: node number n, connection type encoding rule) is used to transform the above mechanical coupling link diagram into a matrix M, where each matrix element... Calculated using the following relation:

[0011] in, This indicates that the i-th component transmits force or displacement to the j-th component. For the join function, Indicates the force transmission coefficient between components. Indicates the displacement transfer ratio; Furthermore, by using a method for calculating the force transmission coefficient (parameters: component contact area a, friction coefficient μ, material elastic modulus E), the ability to... The numerical values ​​are obtained, and a quantitative force mapping matrix is ​​obtained; Furthermore, by using the displacement transfer ratio analysis method (parameters: number of teeth z of the transmission component, pitch p, stroke l), the displacement transfer ratio can be analyzed. The numerical values ​​are obtained, and the final numerical version of the comprehensive force-displacement coupling matrix M is generated. The mechanical coupling link matrix establishment module transforms the results of the previous step into spatial connectivity data indicators, enabling quantifiable modeling of node connection relationships in the causal skeleton diagram and providing morphological constraints for subsequent cross-domain mechanism rule mapping. For example, in a powder packaging machine with a rated capacity of 60 bags per minute, CAD assembly drawings are obtained and the spatial connection sequence of the feeding screw (node ​​number 1), weighing sensor (node ​​number 2), pneumatic gate (node ​​number 3), and vibrating feeder (node ​​number 4) is analyzed. Constraint types are set to three categories: revolute joint, sliding joint, and fixed connection, generating an initial mechanical coupling diagram. The friction coefficient is set. = Component contact area = cm², material elastic modulus = × 5 MPa, the connection relationship is obtained through the force transmission calculation module. Matrix. Let the pitch be... = mm, number of teeth on transmission gears = ,journey = mm, obtained through the displacement transfer ratio module Matrix. and Input connection function Generate a mechanical coupling matrix M, where non-zero elements are as follows: = This indicates that the feed screw has a significant force transmission effect on the weighing sensor. The output matrix M is stored as a two-dimensional structure array, which is used to determine the causal edge direction of the mechanism rule set R in step S1.2, thereby achieving accurate spatial modeling of the connectivity of equipment nodes; S1.2: Based on the law of conservation of mass, rheological hysteresis characteristics and torque balance equations in the powder flow process, the key process physical constraints affecting the filling accuracy are extracted, and a theoretical derivation model of the observable parameter change trend is established based on these constraints; the change direction of each sensor output under a specific fault mode is calculated through this model (such as screw wear leading to a reduction in filling volume, which in turn causes a continuous positive deviation in the weighing residual), forming a mechanism rule set R to guide the orientation of causal edges; Based on the continuous flow of powder materials during the operation of the powder quantitative packaging equipment, the input conditions include the dynamic motion parameters of mechanical components such as the feeding screw, pneumatic gate, and vibrating feeder, the real-time mass output of the weighing sensor, and the time series data of the ambient temperature and humidity sensor, combined with the mechanical coupling link matrix M constructed in the preceding S1.1 as the connection relationship constraint. The mass conservation law criterion method (parameters: instantaneous mass values ​​of material inflow and outflow, equipment operating cycle Δt) is used to construct an equation for the powder mass balance during the packaging cycle and calibrate the flow transfer coefficients of each link; Furthermore, by using the rheological hysteresis characteristic analysis method (parameters: material viscosity coefficient η, particle size distribution function f(d)), the hysteresis response time τ of powder in the feeding screw and discharge channel is quantitatively derived, and the flow rate attenuation factor of each process is obtained. Furthermore, the torque balance equation solution method is adopted (parameter: screw torque). Bearing friction torque Pneumatic gate reaction torque This allows for dynamic equilibrium analysis of the force state in the mechanical transmission chain and generates the influence coefficient of force changes on the actual filling amount. ; The model is constructed using theoretical derivation steps, combining three types of constraint equations—mass conservation, rheological hysteresis, and torque balance—to form a physical model with observable parameter variation trends. For example, the mass conservation part calculates the mass deviation during the packaging cycle using the following formula. :

[0012] in The quality of material inflow during the cycle. For the quality of material outflow; The rheological hysteresis component is calculated using the following formula to determine the velocity attenuation factor. :

[0013] in The viscosity coefficient, The decay time constant; The torque balance section calculates the fill volume influence coefficient using the following formula. :

[0014] in The screw driving torque, For frictional torque, This is the reaction torque of the gate; The model calculates the direction of change in the output of each sensor under specific fault modes, such as screw wear. Decrease leads to the period The continuous positive bias leads to a continuous positive bias trend in the weighing residual. This trend is then injected into the mechanism rule set R and a directional guidance relationship is established with the corresponding node in the mechanical link. The algorithm described above transforms the results of the previous step into rule items contained in the mechanism rule set R. Each rule item describes the drift source, observable parameters and change direction, thereby achieving the physical interpretability of the causal edge orientation process. For example, in a powder packaging device with an operating cycle of 2.0 seconds, the feed screw speed is maintained at 150 rpm, and the screw drive torque is detected. =15 N·m, frictional torque =3 N·m, gate reaction torque =1 N·m. The ambient humidity sensor recorded a relative humidity of 65%, and the powder viscosity coefficient was... =0.004 Pa·s, hysteresis response time =0.8s. The mass of the inflow is calculated using the law of conservation of mass. =0.505kg, outflow mass =0.500kg, then =0.005kg. The rheological hysteresis formula calculates λ≈0.9968, and the torque balance formula calculates β≈0.7333. Based on the model, the weighing residual is determined to be continuously positively biased. The corresponding rule in the mechanism rule set R is recorded as "Drift source: screw wear; observable parameter: weighing residual; direction of change: positive bias," and a confidence level of 'high' is assigned to this rule. This rule will be mapped to the mechanical coupling link matrix M in S1.4 to form a directional causal edge. S1.3: Obtain a list of 12 typical drift sources, including screw wear, increased ambient humidity, and pneumatic response hysteresis. Combine historical maintenance records and FMEA analysis results to assign a unique fault semantic code to each type of drift source. At the same time, define 8 observable parameters, such as weighing residual sequence, motor current spectrum entropy value, and material feeding time variance, as observation variable nodes and uniformly name them 'observable parameter nodes' as terminal response units in the causal skeleton diagram. S1.4: Based on the mechanical coupling link matrix and mechanism rule set, perform cross-domain mapping operation to directionally associate each type of drift source with the observable parameter anomalies it may cause; use directed edges to represent the causal influence path from the drift source node to the observable parameter node, and label the influence type (positive / negative), propagation delay interval and confidence level in the edge attributes, and finally generate a static parameter drift causal skeleton graph with semantic annotation; S1.5: Perform topological consistency verification on the generated static parameter drift causal skeleton graph, use graph traversal algorithm to detect whether there are loops or isolated nodes, and verify the rationality of the critical path through expert experience base; after confirmation, serialize the static parameter drift causal skeleton graph in JSON-LD format and store it in the local configuration database as the initial inference graph structure input for subsequent multi-source evidence fusion and A* reverse tracing algorithm.

[0015] Step S2: In each calibration trigger cycle, four types of heterogeneous time-series data are collected from the weighing module, drive motor encoder, environmental sensor, and vision-assisted positioning module. Sliding window segmentation and zero-mean normalization are performed on the data from each channel to eliminate dimensional differences and generate a standardized time-series segment sequence. Specifically, this includes: S2.1: Based on industrial fieldbus communication protocols (such as EtherCAT or Modbus TCP), four types of heterogeneous time-series data streams with an original sampling frequency of 1kHz are synchronously acquired from the weighing module, drive motor encoder, ambient temperature and humidity sensor and vision-assisted positioning system. Among them, the weighing signal is the mass residual sequence, the motor encoder outputs the angular displacement pulse sequence, the ambient sensor provides the temperature and humidity sliding average, and the vision module returns the timestamp sequence of the powder contour image at the feed port, which serves as the initial input for multi-source data fusion. Based on the industrial fieldbus communication protocol (protocol type: EtherCAT or Modbus TCP), the multi-channel synchronous acquisition command is invoked to realize the acquisition of raw signals from the weighing module, drive motor encoder, ambient temperature and humidity sensor and vision-assisted positioning system. Employs a time-synchronized triggering mechanism (parameter: sampling reference frequency). Hz, trigger delay tolerance (ms) to achieve sampling time alignment of four types of heterogeneous data streams, so as to ensure a one-to-one correspondence between signals collected by different sensors in the time domain; By using a protocol frame parsing algorithm (the field mapping table is based on the equipment manufacturer's protocol description), and field types such as signed integer, floating-point, and byte array, the numerical unpacking and unit conversion of the quality residual sequence of the symmetrical weighing module are realized, and the original code value output by the hardware is mapped to the actual quality residual data stream (unit: kg). By polling the address register and parsing the angular displacement pulse sequence output by the encoder of the drive motor, the pulse-angular displacement conversion formula is used.

[0016] in, This represents the angular displacement of the motor. This represents the number of pulses received within the current sampling period. For encoder pulse resolution, To obtain the transmission ratio of the reduction mechanism, the displacement of the motor is quantitatively acquired; The sensor data fusion interface is used to call the ambient temperature and humidity sampling command. The signal is passed through a first-order moving average filter to calculate the moving average, so as to suppress instantaneous measurement fluctuations and obtain a smooth time series of temperature and humidity signals. The frame acquisition API of the vision-assisted positioning module is used to obtain the powder outline image at the feed port, and the image timestamp extraction function is called to index each frame timestamp sequence with the corresponding image data to form a visual observation data stream arranged based on the time axis. Through the above-mentioned multi-source synchronous acquisition and protocol parsing processing, the weighing residual sequence, motor angular displacement sequence, temperature and humidity sliding average, and visual contour timestamp are jointly encapsulated into a raw sampling data packet to realize the initial input preparation for multi-source data fusion. For example, in a certain powder packaging production line, the weighing module is configured with a sampling frequency of... Hz, the encoder resolution of the drive motor is Pulse / revolution, gear ratio of the reduction mechanism is The environmental sensor's temperature / humidity sampling period is The image acquisition rate of the vision module is s, which is 1 second. Frame. The fieldbus uses the EtherCAT protocol, and the sampling delay is controlled within... Within milliseconds. After executing the above acquisition process, the raw code value output by the weighing module is converted to obtain a continuous mass residual curve; the motor encoder data is converted to an accurate rotation angle time series using the angular displacement conversion formula; the temperature and humidity signals are filtered by moving average to generate a smooth sequence; and the time series of the powder outline at the feed port is formed by the vision module output through timestamp indexing. These four types of heterogeneous data jointly constitute an accurate, time-aligned raw sampling data package, supporting subsequent sliding window segmentation and normalization processing, and significantly improving the consistency and timeliness of multi-source data. S2.2: Perform an equal-length sliding window segmentation operation on each type of raw time-series data stream, with the window length set to 1.5 times the single packaging cycle (i.e., (), step size is 0.5× This generates an overlapping sequence of time segments, ensuring that dynamic transition features at the start and end boundaries of the packaging action are captured, while preserving the continuity of the event context. S2.3: For each time segment within a sliding window, calculate its channel-specific baseline value. For weighing residuals and motor current, use median to estimate the trend term. For temperature and humidity signals, use sliding exponential average filtering. For visual contour area sequences, extract the background baseline through morphological opening operations to eliminate the effects of non-steady-state bias. S2.4: Perform zero-mean and standard deviation normalization on each channel time segment after debiasing. Specifically, subtract the mean of the segment and divide by its standard deviation to uniformly map signals with different physical dimensions (such as kg, rpm, %RH, pixel²) to a dimensionless space, generating a comparable standardized time segment sequence to avoid scale-dominated bias in the subsequent feature extraction process. S2.5: The four types of normalized time series segments are structured and encapsulated according to device ID, timestamp and data source label to generate multi-source time series data packets with metadata annotations, and cached in the local shared memory area as direct input for the next stage to extract the three statistical features of kurtosis, skewness and autocorrelation decay time constant, ensuring data consistency and timeliness.

[0017] like Figure 2 As shown, step S3 involves extracting three statistical features—kurtosis, skewness, and autocorrelation decay time constant—for each type of standardized time series segment to form a multi-dimensional feature vector. Each feature is then mapped to the corresponding observable parameter node in the static parameter drift causal skeleton graph, generating a multi-source observation evidence set with a quantized intensity scalar. Specifically, this includes: S3.1: Based on the standardized time-series segment sequence generated in S2, four types of heterogeneous time-series data segments output by the weighing module, drive motor encoder, environmental sensor, and vision-assisted positioning module are obtained as input objects for feature extraction; using the zero-mean normalized signal after sliding window segmentation, feature bias caused by differences in dimensions and amplitudes between channels is eliminated, and time-series data units in a unified format are generated, providing a consistent input basis for subsequent multi-dimensional feature calculation; Based on the standardized time-series segment sequence generated by S2 and encapsulated by device ID, timestamp, and data source label, the data channel decapsulation method (parameters: source label filtering rules, device ID matching conditions) is used to accurately extract four types of heterogeneous time-series data segments: weighing module, drive motor encoder, environmental sensor, and vision-assisted positioning module. Furthermore, by using a data type matching algorithm (parameters: signal type mapping table, sampling frequency consistency constraint), a unified ontological identification of different physical quantity signals is achieved, and a heterogeneous signal set including mass residual, angular displacement pulse, temperature and humidity sliding mean and powder contour area is obtained; Furthermore, a sliding window segment index reconstruction method (parameters: window start and end index matrix, step size setting) is adopted to align each channel segment under a unified time base and generate multi-channel aligned data blocks with consistent timestamp sequences; Furthermore, by using the zero-mean normalized vector recombination algorithm (parameters: mean μ calculation rule, standard deviation σ calculation rule), the difference in dimensions and amplitude across channels is eliminated, and a dimensionless signal matrix is ​​obtained; Furthermore, a standardized encapsulation method (parameters: channel label set, timestamp index table) is adopted to construct a unified data unit for four types of normalized signals and generate a set of time-series data units with consistent input interfaces. Through the above data decapsulation, type matching, time alignment, normalization and unified encapsulation processing, the multi-source heterogeneous standardized fragments in the previous step are transformed into unified format time series data units that can be directly used for multi-dimensional feature calculation such as kurtosis, skewness and autocorrelation decay time constant, thereby achieving consistency in cross-source feature extraction and improving computational performance. For example, in a powder quantitative packaging production line, a weighing module with a sampling frequency of 1000Hz transmits the residual mass signal, the drive motor encoder outputs an angular displacement sequence of 500 pulses per revolution, environmental sensors record the moving average of temperature (24.6°C) and humidity (61.2%), and a vision positioning module provides the contour area data (unit: pixel²) of the powder discharge port. In this sub-step, signal decapsulation is performed according to the device ID "PKG-LINE-03" and the timestamp range 08:00:00~08:10:00 to obtain the channel segments. Using a signal type mapping table, the residual mass, angular displacement, temperature and humidity, and contour area are identified as a unified input domain. A window length of 1.5 times is used. (in The index is reconstructed with a step size of 1.0s and a time interval of 2.0s to achieve fragment time alignment. The mean and standard deviation are calculated for each fragment, such as the mean of the quality residual fragment. kg, standard deviation kg, zero-mean normalization is performed to obtain a dimensionless matrix. The four normalized signals are then uniformly packaged according to the labels "WEIGHT", "ENCODER", "ENV", and "VISION" to form a time-series data unit with a consistent interface. Subsequently, when performing kurtosis, skewness, and autocorrelation feature extraction in S3.2-S3.4, the calculation process is significantly simplified and the accuracy of cross-channel feature comparison is greatly improved. S3.2: For each type of standardized time-series data unit, perform kurtosis calculation processing and use a fourth-order central moment normalization algorithm to quantify the sharpness of the signal distribution in order to assess whether there are pulse-type disturbances or intermittent fault modes during equipment operation; output the calculated results as the first-dimensional feature parameter characterizing the clustering of abnormal fluctuations to form a kurtosis feature parameter, which is used to support the identification of events such as instantaneous screw jamming or sudden leakage of pneumatic components; Based on the unified format standardized time-series data units generated in S3.1, the signal data corresponding to the weighing module, drive motor encoder, environmental sensor and vision-assisted positioning module are selected as the processing objects for kurtosis calculation; The fourth-order central moment normalization algorithm is used (parameters: sample length n, sample mean). Sample standard deviation This allows for the quantification of the sharpness of signal distribution, thus characterizing the pulse aggregation characteristics of the signal in the time domain; Furthermore, through the calculation formula:

[0018] To achieve standardized calculation of the fourth-order center distance, where For sample data points, To determine the sample size, normalization is used to eliminate the impact of different channel amplitudes on sharpness assessment. Furthermore, the results calculated using the above kurtosis formula are compared with the Gaussian distribution reference value 3 to generate a kurtosis offset index, which is used to determine whether the signal exhibits peak-shaped disturbance behavior. By setting a condition for positive over-threshold kurtosis offset to detect pulse-type disturbance events, and attaching the event occurrence timestamp and corresponding data source label, the annotation output of abnormal fluctuation clustering can be achieved. By using the kurtosis calculation chain, the standardized signal from the previous step is transformed into the first-dimensional kurtosis feature parameter, enabling rapid identification of events such as instantaneous screw jamming or sudden leakage of pneumatic components. For example, in one powder packaging calibration cycle, the standardized time-series data of the weighing residual channel is 512 points long, with a sample mean of... =0.002, standard deviation =0.015, substituted into the kurtosis formula The calculated kurtosis value is 4.6, with an offset of 1.6 from baseline 3, significantly higher than the set threshold of 1.2. This corresponds to an impulse interference event. After mapping to the observable parameter nodes via the causal skeleton graph, e i With an intensity scalar of 0.85, it supports subsequent path confidence updates and reverse causal reasoning, achieving high agility and accurate qualitative identification of anomaly detection; S3.3: Based on the standardized time-series data units obtained in S3.2, skewness calculation processing is performed, and a third-order central moment normalization algorithm is used to measure the asymmetry of the signal distribution in order to identify whether there is a continuous deviation trend in the powder feeding process; the calculated results are output as the second-dimensional feature parameter to form the skewness feature parameter, which is used to reflect early signs of slowly evolving faults such as systematically low filling volume caused by unilateral wear of the feeding screw; S3.4: For the same standardized time-series data unit, calculate the decay time constant of its autocorrelation function, obtain the scale of change of the inertial characteristics of the dynamic response of the signal by fitting an exponential decay model; output the autocorrelation decay time constant as the third-dimensional feature parameter to form an autocorrelation decay feature parameter, which is used to characterize the phenomenon of increased persistence of feeding time-series fluctuations caused by the lag in the action of the pneumatic gate or the slow response of the vibrating feeder. S3.5: Combine the kurtosis, skewness, and autocorrelation decay features extracted from S3.2 to S3.4 into a multidimensional feature vector, and map each feature vector to the corresponding observable parameter node in the graph based on the static parameter drift causal skeleton graph constructed in S1; generate a quantization intensity scalar e according to the preset feature-node mapping rules. i A multi-source observation evidence set ∈ [0,1], where e i This indicates the strength of evidence support on the causal edges associated with the node, serving as the data basis for supporting subsequent path confidence updates and reverse tracing inference.

[0019] like Figure 3 As shown, step S4 involves updating the confidence of each path based on the strength scalar of each piece of evidence in the multi-source observation evidence set and the mechanistic weights and physical uncertainty intervals of their corresponding causal edges, combined with historical evidence decay rules. Then, using an improved A* search algorithm, a high-confidence tracing path satisfying the optimal conditions of the scoring function is traced back on the enhanced causal graph, generating an interpretable corrected decision tree containing at least two causal nodes. Specifically, this includes: S4.1: Based on the predefined set of drift source nodes, observable parameter nodes, and directional causal edges in the static parameter drift causal skeleton graph, obtain the multi-source observation evidence strength scalar corresponding to each causal edge. The evidence strength scalar is generated by mapping the previously extracted kurtosis, skewness, and autocorrelation decay features. Assign initial mechanism weights to each causal edge according to prior knowledge of the equipment's physical mechanism, and determine the width of its corresponding physical uncertainty interval by combining historical maintenance data and simulation analysis. Construct a weighted causal graph using the above three types of parameters, which includes a node set, edge set, mechanism weight vector, and physical uncertainty interval set, as the structured input basis for subsequent dynamic confidence inference, to form an enhanced inference graph structure with physical constraints and credibility assessment capabilities. Based on the defined drift source nodes, observable parameter nodes, and directional causal edge set in the static parameter drift causal skeleton graph, a feature mapping extraction method (parameters: kurtosis feature parameter, skewness feature parameter, autocorrelation decay time constant) is used to obtain the evidence strength scalar corresponding to each causal edge from the multi-source observation evidence set. Furthermore, by using a mechanism weight allocation method (parameters: equipment physical mechanism prior database, node association mechanism model), an initial mechanism weight is set for each causal edge, and a mechanism weight vector data structure is obtained. Furthermore, by using the physical uncertainty interval estimation method (parameters: historical maintenance records, finite element simulation results), the width of the physical uncertainty interval for each causal edge is calculated, and a set of physical uncertainty intervals is generated. Furthermore, by using a weighted causal graph construction algorithm (parameters: node set, edge set, mechanism weight vector, physical uncertainty interval set), the evidence strength scalar, mechanism weight, and physical uncertainty interval are structurally integrated to generate an enhanced reasoning graph structure. The causal edge weights in the path scoring function are initialized using the following formula:

[0020] in, Let the initial weights of the causal edges be... The average effect coefficient prior to the mechanism. It serves as a variance index for the interval of physical uncertainty. By constructing a causal graph, the results of the previous step are transformed into an enhanced inference graph that includes physical constraints and credibility assessment capabilities, thus providing a structured input basis for subsequent dynamic confidence inference. For example, in a powder quantitative packaging production line, the static causal framework diagram includes 12 drift source nodes (such as screw wear, increased ambient humidity, etc.), 8 observable parameter nodes (such as weighing residuals, motor current spectral entropy, etc.), and a corresponding set of oriented causal edges. In one calibration cycle, the kurtosis is 3.8, the skewness is 0.65, and the autocorrelation decay time constant is 2.3s. After mapping, the evidence strength scalar of a certain causal edge is 0.72. Combined with the mechanistic database, the mechanistic weight of this edge is given by the formula... The calculated value is approximately 2.125. Historical maintenance data and simulation analysis indicate that the width of the physical uncertainty interval for this edge is 0.12. By constructing a weighted causal graph algorithm, the node set, edge set, weight vector, and uncertainty set are merged into an enhanced inference causal graph. The graph structure exhibits high robustness to humidity-sensitive drift paths during the subsequent confidence update process in S4.2, significantly improving the accuracy of high-confidence path identification. S4.2: Based on the historical multi-source observation evidence sequence of multiple consecutive calibration cycles within the sliding time window in the weighted causal graph, the evidence strength on each causal edge is time-weighted and aggregated. The effective evidence strength at the current moment is calculated using a power-law decay rule to highlight the dominance of recent observations while preserving the long-term trend memory effect. An enhanced evidence strength vector is output as a dynamic input factor for the path confidence scoring function, thereby generating an evidence support field that integrates spatiotemporal characteristics to support a sensitive response to evolutionary failure modes in the subsequent path optimization process. S4.3: Based on the enhanced evidence strength vector, mechanism weight vector, and physical uncertainty interval set, a path confidence scoring function is constructed. This function quantifies the overall confidence level of the entire tracing path by accumulating the product of the mechanism weight, effective evidence strength, and uncertainty penalty term on each causal edge. The scoring function is then traversed and calculated in the global causal path space to generate the overall confidence score of each candidate reverse tracing path, forming a path candidate set and its corresponding scoring matrix. This serves as the evaluation basis for the improved A* search algorithm, and outputs a path optimization ranking table with physical interpretability to support the accurate identification of high-confidence causal chains. Based on the enhanced evidence strength vector, mechanism weight vector, and set of physical uncertainty intervals, a causal path scoring construction method is adopted (parameter: evidence strength vector e). i Mechanism weight w i Physical uncertainty interval δ i This enables the quantification of the overall confidence level of the entire candidate reverse tracing path; Furthermore, the path cumulative product summation algorithm is used (parameter: set of causal edges on the path). This allows for the calculation of the overall path score and the determination of the confidence level of each candidate path in the current enhanced cause-effect graph. Furthermore, the following scoring function is used to quantify the path confidence level:

[0021] in, The overall confidence score for the path. Mechanism weighting, For the strength of valid evidence, The width of the physical uncertainty interval. The number of causal edges on the path; Furthermore, by using a path exhaustive traversal algorithm (parameter: global causal path space P), the scoring function is fully covered on all paths, and a path candidate set and its corresponding comprehensive scoring matrix are generated. Furthermore, a scoring matrix screening mechanism is employed (parameter: scoring threshold). This enables the optimal ranking of the path candidate set and yields a list of highly confident paths with strong physical interpretability. By constructing a score and performing a traversal calculation, the enhanced evidence strength vector from the previous step is transformed into a comprehensive confidence score index for the path, thereby achieving the goal of ranking the confidence levels of causal paths across the entire domain. For example, in a powder packaging production line, the enhanced evidence strength vector is [0.78, 0.65, 0.81], the mechanism weight vector is [1.2, 0.9, 1.5], and the set of physical uncertainty intervals is [0.04, 0.07, 0.05]. These parameters are then applied to a scoring function. Calculate the contribution value of each causal edge in sequence: First edge The first side is approximately 3.245, the second side is approximately 2.042, and the third side is approximately 3.318. These values ​​are summed to obtain a comprehensive candidate path score of 8.605. During path traversal, the above calculation is repeated for other paths in the global causal path space to form a scoring matrix. A scoring threshold is then set. =6.5, and the paths with scores greater than this threshold are selected into the preferred set. After sorting, the path with a score of 8.605 is selected first for subsequent reverse tracing using the improved A* search algorithm to achieve accurate identification and interpretation of the dominant drift source chain. S4.4: Based on the path score matrix and the current quality deviation as the reverse tracing target, an improved A* search algorithm is executed. During the search process, a heuristic function is introduced to estimate the score of the most likely path from any intermediate node to the root node, and a pruning threshold based on the statistical characteristics of the score distribution is set to remove low-confidence branches. Nodes with high evaluation function values ​​are preferentially expanded in the open list, and the node access status is dynamically maintained until convergence to the optimal high-confidence tracing path that satisfies the comprehensive confidence score not lower than the pruning threshold and contains no less than two causal nodes. The drift source sequence and intermediate evidence links associated with this path are output as the backbone structure of the interpretable correction decision tree to ensure that the diagnostic results have sufficient causal chain support. Based on the path rating matrix and the current quality deviation data input, an improved A search algorithm (parameters: path rating matrix S, target quality deviation Δm, pruning threshold τ) is used to realize a high-confidence path search function with the current deviation as the reverse tracing target; Furthermore, by introducing heuristic function estimation during the search process, the score of the maximum possible path from any intermediate node to the root node is estimated using the following expression: ,in For heuristic estimates, Assign a confidence score sign to a single path from the current node n to the root node r to determine the priority of the search order and dynamically adjust the node expansion strategy in the open list; Furthermore, by setting a pruning threshold based on the statistical characteristics of the scoring distribution, a formula is constructed using variance plus mean: ,in The mean score for the path. To achieve the standard deviation of the score, low-confidence branch paths are eliminated, reducing the invalid search space and improving computational efficiency; Furthermore, by prioritizing expansion of the highest-ranking open list... Valuation points, using formulas ,in For the accumulated path score, To provide heuristic estimation values, and to enable the expansion of pathways based on comprehensive evaluation; Furthermore, by dynamically maintaining a node access status table, expanded nodes are identified to avoid duplicate searches, and all candidate paths converge to a point where the overall confidence score is not lower than a certain threshold. And after including the condition of having no less than two causal nodes, output the optimal high-confidence tracing path; By combining the path score matrix with the quality deviation through the improved A search algorithm, the drift source sequence and intermediate evidence link data are transformed into the backbone structure of the interpretable correction decision tree. For example, in one operation cycle of a powder quantitative packaging equipment, the maximum score of the path scoring matrix is ​​95, the mean is 78, the standard deviation is 5, and the pruning threshold is... The current quality deviation Δm is 0.45 kg. The maximum path score from the candidate node to the root node, estimated using a heuristic function, is 88. Combined with the accumulated score... Comprehensive evaluation value Paths higher than other nodes in the open list are expanded first. During the expansion process, paths below a threshold (e.g., scores of 80) are pruned, and the final search result converges to a path containing three causal nodes: "screw wear → decreased filling efficiency → positive metering bias". The corresponding intermediate evidence link is formed by a combination of weighing residual kurtosis features and motor current skewness features. This search process reduces computation time while ensuring that the diagnostic results have a complete causal chain to support them. The output path is directly used for decision tree construction and enters the downstream parameter correction instruction generation module. S4.5: Based on the node sequence in the optimal high-confidence tracing path, it is transformed into a hierarchical interpretable correction decision tree structure: with the current quality deviation as the root node, the activated drift source types are used as intermediate nodes in sequence, and their corresponding evidence support percentages are labeled; the terminal leaf nodes generate specific parameter compensation suggestions, and attach the original data index number, feature calculation formula label, and confidence level identifier; the output is a structured reasoning result with a complete causal chain and supporting evidence, which is used by the subsequent instruction conversion module to generate an executable correction instruction package, realizing the semantic connection from abstract diagnostic conclusions to specific control actions; Based on the node sequence of the optimal high-confidence tracing path, a hierarchical tree structure generation method is adopted (parameters: node sequence, causal type label, confidence percentage) to realize the structural mapping of the current quality deviation as the root node of the decision tree. Furthermore, by using the node semantic binding method (parameters: drift source type, evidence support), the activated drift source type is embedded into the intermediate level of the tree step by step, and the evidence support percentage is marked in the corresponding node attributes to obtain an intermediate node set with diagnostic interpretability. Furthermore, by using a parameter compensation suggestion generation algorithm (parameters: end node diagnostic conclusion, control parameter dictionary), specific process parameter suggestions for the end leaf node are generated, and the suggestions are bound to controllable parameter types such as filling volume compensation coefficient or material feeding delay threshold, resulting in an end suggestion set with direct execution capability. Furthermore, an original data index referencing mechanism (parameters: node-associated data fragment ID, feature formula hash value) is adopted to attach an original data index number and feature calculation formula identifier to each terminal leaf node, ensuring that the diagnostic basis can be traced back; Furthermore, the confidence level labeling algorithm (parameters: path comprehensive score value, level classification threshold) is executed to label each node to form a level label, thus forming a complete causal chain visualization data package; Through the above multi-level mapping and attribute binding processing method, the optimal high-confidence path result of the previous step is transformed into a structured reasoning tree containing the root node quality deviation, intermediate causal nodes, end compensation suggestions and their supporting data, so as to realize the semantic connection from diagnosis to control action. For example, in a powder packaging equipment, the optimal high-confidence traceability path node sequence is {mass deviation Δm = 0.025 kg, drift source A: increased screw clearance (evidence support 0.84), drift source B: decreased filling coefficient (evidence support 0.78)}, and the corresponding end parameter compensation suggestion is to increase the filling volume compensation coefficient by 2.7%. First, a hierarchical tree structure generation method is used, with the Δm node as the root node, and the node attributes set to... Secondly, drift source A and drift source B are sequentially added to intermediate nodes using a node semantic binding method, with evidence support scores of 0.84 and 0.78 respectively. Subsequently, a parameter compensation suggestion generation algorithm is used to map the terminal suggestions to control parameters. The correction amount of 2.7% was calculated using the path comprehensive scoring formula. Then, the weighing residual segment ID=CR_145 and the feature formula hash value H_F123 were bound to the terminal leaf node through the original data indexing reference mechanism. Finally, the path was marked as "high confidence" using a confidence level identification algorithm. The structured inference tree obtained after executing the above chain successfully drove the controller to adjust the filling volume compensation coefficient in practical applications, significantly reducing the quality deviation of the equipment in the subsequent 10 packaging cycles, achieving a seamless transition from diagnosis to execution.

[0022] Step S5: The root node error, intermediate causal links, and terminal parameter compensation suggestions in the interpretable correction decision tree are converted into a structured correction instruction package with weighted labels and original data references. The compensation suggestions explicitly point to controllable process parameters such as the filling volume compensation coefficient or the material feeding delay threshold. Specifically, this includes: S5.1: Based on the root node error Δm and the confidence weight w of each node in the intermediate causal link, the decision tree output is corrected. i and the corresponding strength of observational evidence e i We construct a metadata header field that includes error tracing path identifiers, key feature tuple indexes, and original data timestamp references to form the basic framework of a structured correction instruction package, which will be used to carry parameter compensation suggestions and their supporting basis in the future. Based on the root node error Δm of the interpretable corrected decision tree output, and the confidence weights of each node in the intermediate causal link. and the corresponding strength of observational evidence The path information parsing method (parameters: node sequence, edge attribute set) is used to extract error source tracing path identifiers from the corrected decision tree. Furthermore, by using a feature index mapping algorithm (parameters: node feature calculation formula label, feature hash table), the key feature tuple index is generated, and a feature index dataset that can be used for instruction packet association is obtained. Furthermore, by using a timestamp association method (parameters: original data fragment ID, sampling time, event index table), the binding between the original data reference and the traceability path identifier is realized, and a timestamp reference field is generated to accurately locate the supporting data source when executing instructions; Furthermore, a metadata encapsulation processing method (parameters: path identifier, feature tuple index, timestamp reference) is adopted to construct the metadata header field and generate a structured description to ensure that it can carry parameter compensation suggestions and supporting evidence; The metadata encapsulation algorithm transforms the parsing results of the previous step into the basic framework of the correction instruction package, achieving semantic connection and format consistency between metadata and subsequent control parameter fields. For example, in a self-calibration cycle of quantitative powder packaging, the error Δm of the root node output by the decision tree is... (Unit: g), the intermediate causal link contains three drift source nodes, whose confidence weights are respectively = , = , = corresponding strength of observational evidence = , = , = The path information parsing method was used to parse the source path identifier as "PTH-A17". The node feature formula label was mapped to the index set {IDX-022, IDX-145, IDX-311} using the feature index mapping algorithm. Then, the original data segment timestamps {2024-06-21T14:35:22Z, 2024-06-21T14:35:25Z, 2024-06-21T14:35:27Z} were bound using the timestamp association method. Metadata encapsulation integrates the aforementioned identifiers, index sets, and timestamps into a header field structure: {"path_id":"PTH-A17","feature_idx":["IDX-022","IDX-145","IDX-311"],"timestamp":["2024-06-21T14:35:22Z","2024-06-21T14:35:25Z","2024-06-21T14:35:27Z"]}. This structure serves as the carrier for parameter compensation suggestion binding in subsequent S5.2, realizing the semanticization and precision of the control instruction framework, and maintaining the consistency and traceability of instruction parsing during the execution phase; S5.2: For the parameter compensation suggestions generated at the end of the decision tree, identify the specific controllable process parameter type they point to—if the suggestion involves feeding volume deviation, map it to the filling volume compensation coefficient; if the suggestion is associated with material feeding dynamic lag, parse it as the material feeding delay threshold; perform semantic binding through the preset process parameter dictionary table to generate parameter compensation suggestion type labels, ensuring that the control commands are consistent with the equipment execution layer interface; S5.3: Calculate the correction magnitude for each parameter compensation suggestion: based on the gain model of the influence of the dominant drift source in the causal chain. (e.g., the empirical calibration curve of screw wear level → filling efficiency reduction rate), combined with the current residual Δm and the path comprehensive confidence score, a linear proportional allocation algorithm is executed to obtain... or The specific adjustment amount generates a numerical control variable with unit labels; Based on the controllable process parameter type labels output by S5.2 and the error tracing path metadata constructed in S5.1, the input objects include the current quality residual. Path-wide confidence score and the influence-gain model of the dominant drift source in the causal chain. Parameter set; Model mapping method is used (parameter: influence gain model) This allows the dominant drift source level to be mapped to the corresponding process efficiency change rate, and generates an efficiency decay ratio coefficient matrix as the model weight for the correction amplitude calculation. Furthermore, through the residual coupling calculation method (parameter: current residual) The efficiency decay ratio matrix is ​​used to decompose the quality residual into the target parameter influence domain according to the mechanism correlation, and obtain the residual component scalars. Furthermore, a confidence-weighted calculation method is used (parameter: path comprehensive confidence score). This allows the residual components to be multiplied by the confidence field to obtain a weighted residual effect value, which is used for parameter magnitude allocation. Furthermore, a linear proportional allocation algorithm (parameters: weighted residual effect value, target parameter type label) is used to proportionally allocate the weighted residual effect value to the filling volume compensation coefficient. Or material feeding delay threshold And generate numerical control variables with unit labels; Through the above algorithm chain, the result of the previous step is converted into a precise numerical correction amplitude, thus realizing the engineering feasibility of process parameter adjustment. For example, the gain model of the effect on feed screw wear level 3. Set as Current quality residual for kg, path comprehensive confidence score for The correlation coefficient of the efficiency decay ratio matrix obtained by using the model mapping method is: The scalar of the residual components is obtained through residual coupling calculation as follows: kg, weighted residual effect value kg. This effect value is allocated to the filling volume compensation coefficient using a linear proportional allocation algorithm. The specific adjustment amount needs to be determined. L, in liters, outputs numerical control variables and binds them to the corresponding registers, thus significantly improving fill accuracy; S5.4: Package and integrate the metadata header fields, parameter compensation suggestion type label, numerical control variables, original data fragment IDs that support the suggestion (from weighing residual waveforms, motor current spectrum, etc.) and the corresponding feature extraction formula hash values, and serialize them into a structured correction instruction package using JSON-LD format to ensure that the instruction content can be jointly parsed by the downstream rendering engine and controller. S5.5: Output structured correction instruction packets to a dual-channel distribution queue: one channel is sent to a lightweight WebGL rendering engine for visualization and source map generation, and the other channel is cached and sent to the PLC control bus after confirmation; at the same time, the instruction packet summary fingerprint is recorded in the local log to support subsequent operation auditing and closed-loop traceability.

[0023] Step S6: Before executing the correction instructions, the structured correction instruction package is input into the lightweight WebGL rendering engine to generate an interactive SVG format dynamic source map, which is then projected onto the human-machine interface in real time, allowing operators to view the supporting evidence waveforms, feature calculation basis, and recommended action prompts for each node. Specifically, this includes: S6.1: Based on the root node error, intermediate causal links and end parameter compensation suggestions contained in the structured correction instruction package, extract the original data reference identifiers (such as weighing residual segment ID, motor current spectrum segment index), feature calculation formula code blocks and confidence label values ​​corresponding to each node, and generate a set of metadata of graph elements with semantic tags, which serves as the basic input for subsequent SVG graphic element mapping. Based on the root node error, intermediate causal links, and terminal parameter compensation suggestions in the structured correction instruction package, a field parsing method (parameters: JSON-LD format parsing rules, semantic tag mapping table) is used to separate the core information and identify the attributes of each causal node. Furthermore, by using the original data index parsing algorithm (parameters: timestamp alignment rules, fragment ID mapping matrix), the accurate extraction of original data reference identifiers such as symmetrical residual fragment ID and motor current spectrum segment index is achieved, and the original data location results corresponding to the node order are obtained. Furthermore, by using the formula code block extraction and parsing method (parameters: feature calculation formula hash value mapping table, formula entityization rules), the feature calculation formula code blocks supporting the evidence of each node are restored and bound, ensuring that the formula text and node semantics are accurately associated. Furthermore, using a confidence labeling generation algorithm (parameters: scoring matrix, normalized interval [0,1]), the calculation results of evidence strength and mechanism weight are combined to generate the confidence label value corresponding to each causal node; where confidence... The calculation formula can be expressed as:

[0024] in, To correspond to the mechanistic weights of the causal edges, As a scalar of the strength of evidence, The number of features associated with a node; Furthermore, a semantic tag assignment method (parameters: node type dictionary, attribute mapping table) is adopted to process the drift source name, observable parameter name and its confidence label value into a set of graph element metadata with semantic tags, which can be directly called by the subsequent SVG graphic element mapping module; By combining the above-mentioned field parsing, index extraction, formula restoration, numerical calculation and label assignment processing methods, the correction instruction package generated in the previous step is transformed into a complete, structured and semantically consistent set of map element metadata, thereby achieving a highly consistent input effect in the dynamic traceability map generation process. For example, in a self-calibration scenario of a powder quantitative packaging equipment, the structured calibration instruction package contains a root node error of... kg, the two nodes in the intermediate causal chain correspond to the drift sources "increased feed screw clearance" and "low filling volume compensation coefficient" respectively, and the end compensation suggestion is to adjust. to The parsing algorithm extracts the weighing residual segment IDs as #WR_10582 and #WR_10585, and the motor current spectrum segment index as #MCF_884. The formula code block is restored to kurtosis calculation through hash mapping. The confidence assignment module calculates the confidence scores of the two nodes as follows: and The semantic tag assignment module encodes the drift source, parameter name, and confidence level into readable metadata items. The final generated set of graph element metadata is input into the SVG mapping stage, significantly improving the traceability and display accuracy of causal chain evidence in subsequent rendering. S6.2: For each causal node and directional edge in the metadata set of the graph elements, use a predefined graph layout algorithm (based on the force-oriented model) to perform two-dimensional coordinate allocation, calculate the spatial position distribution of each node in the HMI display area, and generate a node coordinate matrix with physical constraints to optimize the readability of the graph and avoid visual overlap. S6.3: Based on the node coordinate matrix and graph element metadata, the vector drawing interface of the lightweight WebGL rendering engine is called to build a dynamic source map in SVG format layer by layer: First, draw causal edges with color gradient encoding (the color depth corresponds to the path confidence), then generate a circular node container, fill the text label inside the node (including the drift source name, observable parameter name and confidence percentage), and finally embed a clickable interactive control. Based on the node coordinate matrix and graph element metadata, a vector drawing interface of a lightweight WebGL rendering engine (parameters: supports SVG output, color gradient encoding range [0,255], coordinate mapping resolution 1920×1080) is used to realize the function of drawing causal edges layer by layer. Furthermore, through a color gradient encoding algorithm (parameters: path confidence value q∈[0,1], color mapping function...), This achieves a linear correspondence between the color depth of causal edges and the path confidence value, and obtains the edge set rendering result with color gradient attributes; Furthermore, through a node container generation algorithm (parameters: node radius r=16px, border thickness 2px, fill color RGBA(230,230,230,1)), a circular container is constructed for all causal nodes, and a set of node primitives is generated to hold text labels and interactive controls. Furthermore, through the text label filling process (parameters: font Family=Arial, font Size=12pt, alignment=center), the drift source name, observable parameter name, and confidence percentage are embedded into the node container, and node elements with text attributes are generated. Furthermore, by using the interactive control binding method (parameters: event type onclick, id binding rule node_i), clickable interactive controls can be attached to each node container, and a list of interactive nodes that can trigger events can be generated. Through the above-mentioned drawing and interactive component construction process, the node spatial coordinates and metadata annotations of the previous step are transformed into a dynamic tracing map in SVG format, realizing the visualization and interactivity of the causal chain in the HMI interface. For example, in a scenario involving parameter drift diagnosis of a powder quantitative packaging equipment, the node coordinate matrix is ​​a 6×2 two-dimensional coordinate array with a maximum coordinate range of [0, 512]. The input graph element metadata contains 6 causal nodes and 5 directional edges, with node confidence percentages distributed as {0.92, 0.86, 0.75, 0.80, 0.78, 0.83}. A color mapping function is used. Each edge is color-coded. When the path confidence q is 0.92, R=20, G=234, B=128 are calculated, resulting in a light green color. For a lower confidence q=0.75, the corresponding color is yellowish-green (R=64, G=191, B=128). A node container with a radius of 16px is generated and filled with a background color of 230 grayscale. The drift source name and confidence percentage in Arial / 12pt font are embedded in the center of the container, such as the node 1 label being "Screw Wear / 92%". An onclick event is bound to each node, with the ID naming rule being node_i (i being the node number). Clicking triggers a waveform thumbnail and expands the formula details. The image is rendered using a WebGL rendering engine, outputting an SVG atlas with a resolution of 1920×1080. Edge anti-aliasing is significantly improved, node text remains sharp on high-DPI displays, and the interactive click response latency is no more than 15ms, demonstrating that this rendering step significantly improves the smoothness of interactive operations and system response speed while maintaining readability. S6.4: Embed a waveform preview thumbnail inside the SVG graph node container. The thumbnail is generated by performing local mean downsampling and amplitude normalization on the standardized time segment pointed to by the original data reference identifier, forming a visual snapshot of supporting evidence, which is used to intuitively show the weighing residual fluctuation or motor current change trend corresponding to the node. S6.5: The rendered SVG format dynamic traceability map is pushed to the designated display area of ​​the HMI interface in real time via the WebSocket protocol, and a front-end event listener is bound to it. When the operator clicks on any node, a floating panel can be expanded to display its complete feature calculation formula, a list of historical similar case numbers, and the system's recommended reset action prompt, so as to realize human-machine collaborative decision support.

[0025] Step S7: If the operator confirms the parameter correction scheme guided by the dynamic traceability map on the human-machine interface, a structured correction instruction package is sent to the controller to drive the actuator to automatically adjust the target process parameters, completing the closed-loop self-calibration action. Specifically, this includes: S7.1: Based on the operator's interactive confirmation signal on the dynamic traceability map in the human-machine interface, obtain the user authorization status identifier; when the identifier is detected as 'confirmed', trigger the activation process of the structured correction instruction package to ensure that all subsequent control actions are based on manual review and approval, and ensure the system's operational security and accountability. S7.2: Based on the structured correction instruction package generated in S5, parse the end parameter compensation suggestion field contained therein, extract the name of the target process parameter to be adjusted and its recommended correction amount; map the target parameter to the corresponding register address table entry in the equipment controller, generate a parameter update list with address index, and use it as the data input basis for actuator control; The structured correction instruction package is parsed using a field parsing method (parameters: JSON-LD format, field type mapping table) to achieve item-by-item reading and type identification of the end parameter compensation suggestion fields; Furthermore, by using the semantic binding method for process parameters (parameters: process parameter dictionary table, semantic tag mapping rules), the name of the target process parameter that needs to be adjusted is accurately extracted, and a list of control variable names that are consistent with the executable parameter identifiers of the equipment control layer is obtained. Furthermore, a correction value parsing algorithm (parameters: recommended correction value, unit label, and numerical validity verification rules) is adopted to separate the recommended correction value in the compensation proposal and bind the unit, and generate a numerical correction variable that can be directly input into the controller. Furthermore, by using the register mapping method (parameters: device controller register address table, parameter-address mapping matrix), a one-to-one correspondence is established between the target process parameter name and its control register address, and a parameter update entry containing a triplet of register address, control variable name, and correction value is generated. A parameter update list construction algorithm (parameters: register address index sequence, update entry set) is adopted to realize the structured encapsulation of multi-parameter batch update lists and obtain a data input list that can be directly controlled by the actuator; Through the above mapping and parsing process, the compensation suggestion result of the previous step is transformed into a parameter update list with address index, realizing a seamless link from diagnostic decision to control execution, and ensuring the accurate transmission and traceable recording of process parameter correction actions. For example, during a self-calibration process of a powder quantitative packaging production line, the end-compensation suggestion field of the structured calibration instruction package includes two items: a filling volume compensation coefficient and a feeding delay threshold. The suggested correction for the filling volume compensation coefficient is 2.7%, in dimensionless proportionality coefficients, and the suggested correction for the feeding delay threshold is 0.35 seconds, in seconds. The field parsing method reads the compensation suggestion field from the JSON-LD format instruction package, and through semantic binding of the process parameter dictionary, maps the filling volume compensation coefficient to the controller variable name "KV_FILL" and the feeding delay threshold to the controller variable name "TD_DELAY". The correction amount parsing algorithm converts 2.7% into... The floating-point value, converting 0.35 seconds to The high-precision values ​​for seconds are all accompanied by unit metadata tags. The register mapping method calls the controller register address table, matching "KV_FILL" to address 0x1A3C and "TD_DELAY" to address 0x1A4F, forming update entries (0x1A3C, KV_FILL, 0.027) and (0x1A4F, TD_DELAY, 0.35). The parameter update list construction algorithm sorts the two update entries according to the address index sequence and encapsulates and outputs them, generating a parameter update list for the actuator to control. It has been verified that the feeding screw drive unit and pneumatic gate in the production line operate stably according to the new parameter settings in the next packaging cycle, and the weighing residual is significantly reduced to the target range. S7.3: Establish a real-time communication channel with the lower-level PLC controller using industrial control protocols (such as Modbus TCP), encapsulate the parameter update list into a standardized write instruction frame; perform parameter write operations based on a periodic synchronization mechanism to achieve accurate assignment of key controllable parameters such as filling volume compensation coefficient or material feeding delay threshold, and generate controlled parameter update instances. S7.4: After the parameters are written, a start synchronization pulse signal is sent to the actuator (such as a servo motor-driven feeding screw or a pneumatic gate solenoid valve) to trigger a new round of standard packaging cycle; by collecting the actual weighing residual data of the first three products after execution, the deviation convergence rate index after calibration is calculated, and a calibration effect feedback dataset is generated to verify the effectiveness of closed-loop regulation. S7.5: Package the execution timestamp of this calibration action, the original error amount, the corrected parameter items, the comparison results of the residuals before and after calibration, and the operator ID into a complete calibration event record, and push it to the write queue of the local time series knowledge graph database as the input source for the construction of RDF triples in S8, supporting subsequent model credibility iteration and historical case retrieval.

[0026] Step S8: Record the dynamic traceability graph structure, user interaction logs, and residual changes before and after calibration during this calibration process. Store these data in the local time-series knowledge graph database as RDF triples. Dynamically update the prior weights of relevant edges in the causal skeleton graph based on human feedback to achieve iterative optimization of the inference model's credibility. Specifically, this includes: S8.1: Obtain the dynamic traceability graph structure data generated during this calibration process, including the root node error, intermediate causal link set and end parameter compensation suggestions, and use the graph serialization protocol to convert it into a raw RDF triple stream containing a subject-predicate-object structure, as the basic input for knowledge storage; S8.2: Collect user interaction behavior logs in the human-machine interface, including operator confirmation / rejection actions of recommended correction schemes, frequency of node expansion and viewing, duration of evidence waveform playback and other interaction event sequences. Generate corresponding RDF assertion sets based on event semantic mapping rules, where the 'operator rejection' event is marked as the causal edge confidence reduction trigger condition. S8.3: Calculate the mean offset of the weighing residuals of 50 consecutive packages before and after calibration and the mean offset of the weighing residuals after calibration. Then bind the change of this performance index to the RDF resource node of the corresponding calibration instance to form the observation fact of 'calibration effectiveness assessment', which serves as an external supervision signal for subsequent weight updates. S8.4: The above-mentioned dynamic traceability graph RDF triples, user interaction RDF assertions and residual performance change facts are spatiotemporally aligned and namespace normalized, merged into a knowledge snapshot of the complete calibration event, and written into a dedicated graph partition indexed by timestamp in the local time series knowledge graph database, supporting joint queries based on SPARQL time range and device ID. S8.5: Determine if there is a need for weight adjustment based on the latest manual feedback results: If the calibration scheme associated with a causal edge is rejected by the operator three times in a row and there is an alternative high-confidence path, then call the Bayesian update mechanism to adjust the prior weight w of that edge. i Perform a decay operation and synchronize the updated weight values ​​to the static parameter drift causal skeleton graph G0 to complete the credibility iterative optimization of the inference model.

[0027] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0028] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A self-diagnosis and self-calibration method for long-term drift of quantitative packaging accuracy of powder, characterized in that, Includes the following steps: S1: Based on the mechanical structure topology and process physical constraints of powder quantitative packaging equipment, construct a causal skeleton diagram of static parameter drift; S2: Collect multi-source heterogeneous time-series data in each calibration trigger cycle, and perform sliding window segmentation and zero-mean normalization on the data of each channel to generate a standardized time-series segment sequence; S3: For each type of standardized time series segment, extract statistical features to form a multi-dimensional feature vector, map each feature to the corresponding observable parameter node in the static parameter drift causal skeleton graph, and generate a multi-source observation evidence set. S4: Based on the strength scalar of each piece of evidence in the multi-source observation evidence set and the mechanism weight and physical uncertainty interval of its corresponding causal edge, update the confidence of each path in combination with the historical evidence attenuation rule, and trace back the high-confidence tracing path that satisfies the optimal condition of the scoring function on the enhanced causal graph to generate an interpretable correction decision tree. S5: Convert the root node error, intermediate causal links, and end parameter compensation suggestions in the interpretable correction decision tree into a structured correction instruction package; S6: Before executing the correction instructions, the structured correction instruction package is input into the WebGL rendering engine to generate a dynamic source map and project it onto the human-computer interface in real time. S7: If the operator confirms the parameter correction scheme guided by the dynamic traceability map on the human-machine interface, the operator sends the structured correction instruction package to the controller to drive the actuator to automatically adjust the target process parameters.

2. The self-diagnosis and self-calibration method for long-term drift of powder quantitative packaging accuracy according to claim 1, characterized in that, The process following step S7 also includes: S8: Record the dynamic traceability graph structure, user interaction behavior logs, and residual change data before and after calibration during this calibration process. Store the data in the local time-series knowledge graph database in the form of RDF triples. Update the prior weights of relevant edges in the causal skeleton graph dynamically based on human feedback results to achieve iterative optimization of the credibility of the inference model.

3. The self-diagnosis and self-calibration method for long-term drift of powder quantitative packaging accuracy according to claim 1, characterized in that, The static parameter drift causal skeleton graph consists of drift source nodes, observable parameter nodes, and directional causal edges.

4. The self-diagnosis and self-calibration method for long-term drift of powder quantitative packaging accuracy according to claim 1, characterized in that, The acquisition of multi-source heterogeneous time-series data specifically involves: Based on the industrial fieldbus communication protocol, four types of heterogeneous time-series data streams with an original sampling frequency of 1kHz are synchronously acquired from the weighing module, the drive motor encoder, the ambient temperature and humidity sensor, and the vision-assisted positioning system. The weighing module outputs a mass residual sequence, the drive motor encoder outputs an angular displacement pulse sequence, the ambient temperature and humidity sensor provides a sliding average of temperature and humidity, and the vision-assisted positioning system returns a timestamp sequence of the powder contour image at the feed port.

5. The self-diagnosis and self-calibration method for long-term drift of powder quantitative packaging accuracy according to claim 1, characterized in that, Step S3 specifically includes: Based on the standardized time series sequence generated in S2, the zero-mean normalized signal after sliding window segmentation is used to generate standardized time series data units in a unified format. For the standardized time-series data unit, kurtosis calculation processing is performed, and the calculated result is output as the first-dimensional feature parameter characterizing the clustering of abnormal fluctuations, forming the kurtosis feature parameter; For the standardized time-series data unit, skewness calculation processing is performed, and the calculated result is output as the second-dimensional feature parameter to form the skewness feature parameter; For the standardized time-series data unit, the decay time constant of its autocorrelation function is calculated, and the decay time constant is output as the third-dimensional feature parameter to form the autocorrelation decay feature parameter. The kurtosis feature parameter, the skewness feature parameter, and the autocorrelation decay feature parameter are combined into a multi-dimensional feature vector. Based on the static parameter drift causal skeleton graph constructed in step S1, each feature vector is mapped to the corresponding observable parameter node in the graph. According to the preset feature-node mapping rules, a multi-source observation evidence set is generated.

6. The self-diagnosis and self-calibration method for long-term drift of powder quantitative packaging accuracy according to claim 5, characterized in that, The multi-source observation evidence set includes the evidence support strength on the causal edges associated with the observable parameter nodes.

7. The self-diagnosis and self-calibration method for long-term drift of powder quantitative packaging accuracy according to claim 1, characterized in that, Step S4 specifically includes: Based on the predefined drift source nodes, observable parameter nodes, and directional causal edge set in the static parameter drift causal skeleton graph, the multi-source observation evidence strength scalar corresponding to each causal edge is obtained. An initial mechanism weight is assigned to each causal edge according to the prior knowledge of the equipment physical mechanism. The width of the corresponding physical uncertainty interval is determined by combining historical maintenance data and simulation analysis. A weighted causal graph is constructed using the multi-source observation evidence strength scalar, the initial mechanism weight, and the width of the physical uncertainty interval. Based on the historical multi-source observation evidence sequence of multiple consecutive calibration cycles within the sliding time window in the weighted causal graph, the evidence strength on each causal edge is subjected to time-series weighted aggregation processing to calculate the effective evidence strength at the current moment and output the enhanced evidence strength vector. Based on the enhanced evidence strength vector, mechanism weight vector, and physical uncertainty interval set, a path confidence scoring function is constructed. The path confidence scoring function is then traversed and calculated in the global causal path space to generate a comprehensive confidence score for each candidate reverse tracing path, forming a path candidate set and its corresponding scoring matrix. Based on the path score matrix and the current quality deviation as the reverse tracing target, an improved A* search algorithm is executed. During the search process, a heuristic function is introduced to estimate the maximum possible path score from any intermediate node to the root node, and a pruning threshold based on the statistical characteristics of the score distribution is set to remove low-confidence branches. Nodes with high evaluation function values ​​are preferentially expanded in the open list, and the node access status is dynamically maintained until convergence to the optimal high-confidence tracing path that satisfies the comprehensive confidence score not lower than the pruning threshold and contains no less than two causal nodes. The drift source sequence and intermediate evidence link associated with the optimal high-confidence tracing path are output. Based on the node sequence in the optimal high-confidence tracing path, it is transformed into a hierarchical interpretable and correctable decision tree structure, and a structured reasoning result with a complete causal chain and supporting evidence is output.

8. The self-diagnosis and self-calibration method for long-term drift of powder quantitative packaging accuracy according to claim 7, characterized in that, The node sequence in the optimal high-confidence tracing path is transformed into a hierarchical interpretable correction decision tree structure as follows: the current quality deviation is taken as the root node, the activated drift source types are taken as intermediate nodes in sequence, and their corresponding evidence support percentages are labeled; the terminal leaf nodes generate specific parameter compensation suggestions, and attach the original data index number, feature calculation formula label, and confidence level identifier.

9. The self-diagnosis and self-calibration method for long-term drift of powder quantitative packaging accuracy according to claim 1, characterized in that, Step S5 specifically includes: Based on the root node error of the interpretable corrected decision tree output, the confidence weight of each node in the intermediate causal link, and the corresponding observation evidence strength, a metadata header field is constructed. For the parameter compensation suggestions generated at the end of the decision tree, the specific controllable process parameter type they point to is identified. If the suggestion involves feeding volume deviation, it is mapped to the filling volume compensation coefficient; if the suggestion is associated with material feeding dynamic lag, it is parsed as the material feeding delay threshold; and semantic binding is performed through a preset process parameter dictionary table to generate parameter compensation suggestion type labels. For each parameter compensation suggestion, calculate its correction magnitude. Based on the influence gain model of the dominant drift source in the causal chain, and combined with the current residual and path comprehensive confidence score, execute the linear proportional allocation algorithm to obtain the specific adjustment amount of the filling volume compensation coefficient or the material feeding delay threshold, and generate a numerical control variable with unit label. The metadata header fields, the parameter compensation suggestion type label, the numerical control variables, the original data fragment IDs supporting the parameter compensation suggestions, and the corresponding feature extraction formula hash values ​​are packaged and integrated, and serialized into a structured correction instruction package using JSON-LD format.

10. The self-diagnosis and self-calibration method for long-term drift of powder quantitative packaging accuracy according to claim 9, characterized in that, The metadata header fields include an error tracing path identifier, a key feature tuple index, and a reference to the original data timestamp.

Images