Intelligent optimization method for packaging parameters of packaging production line

Abstract

The invention discloses an intelligent optimization method for packaging parameters of a packaging production line, and the method comprises the steps: building a minute-level, hour-level and batch-level multi-scale original data set based on the multi-source data, such as sealing temperature, operation speed, tension control, energy consumption index, environment temperature and humidity, and carrying out the standardization and denoising of the original data set, basic change modes of the parameters under different time scales are extracted; through a self-supervised learning multi-scale feature encoder, decoupling of parameter change features is realized, and then through functional semantic mapping, strong correlation between parameters and process targets such as quality stability, energy consumption efficiency, equipment life and the like is established; according to the method, a conflict detection algorithm on a time scale and a function target dimension is fused, a conflict type, an influence path and a contribution degree sequence are output, a dynamic parameter adjustment or target priority reconstruction suggestion is proposed according to conflict intensity, a strategy can be dynamically optimized, accurate identification and efficient intervention of a multi-target conflict of a production line are realized, and the method is suitable for large-scale popularization and application. And the self-adaptive optimization and production intelligence level of the system is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent manufacturing and industrial process optimization, and in particular to a packaging parameter intelligent optimization method for a packaging production line. BACKGROUND

[0002] Currently, in the field of intelligent manufacturing and industrial process optimization, a series of parameter collaborative optimization and conflict management technical solutions have emerged for the multi-objective optimization problem of packaging production lines. The mainstream multi-objective optimization techniques mainly include parameter weighted fusion method, offset function modeling, strategy balance model based on game theory, and causal graph reasoning, etc. These methods usually rely on unified modeling of different process objectives through weighted or mapping functions, and balance and coordination between objectives are achieved by adjusting parameter weights or applying compensation coefficients. For example, in actual production lines, parameter weight optimization algorithms are commonly used, in which different weights are assigned to key parameters (such as temperature, speed, tension, etc.) based on expert experience or historical data, and the overall conflict is minimized by minimizing the weighted objective function. However, such techniques often focus on global optimization and empirical parameter tuning, and lack explainable fine-grained modeling of the nature of parameter conflicts, dynamic transmission paths, and their local influence mechanisms. However, the current mainstream technology still has the following limitations: Firstly, existing parameter conflict modeling methods mostly abstract conflicts as black-box weighted coefficients or offset factors, making it difficult to specifically reveal which parameter, under what working conditions, and with what dynamic behavior, causes objective conflicts. This means that once the production line experiences abnormalities such as decreased sealing strength or increased energy consumption, it is difficult to pinpoint the root cause of the problem relying solely on statistical correlations or empirical rules, making it difficult to support targeted human-machine collaboration or automated strategy adjustment; Secondly, traditional multi-objective optimization mostly focuses on "global balance", and lacks fine-grained analysis of the multi-scale dynamic characteristics of parameter conflicts, such as local instability under rapid disturbance and the nature of mutual exclusion between process objectives. As a result, the system can only find that "there is conflict between objectives", but cannot reveal the specific transmission path of the conflict and its triggering and evolution mechanism over time / function. This limits the implementation of fine-grained intelligent applications such as fault early warning, equipment health management, and proactive prevention of production abnormalities; Thirdly, existing technologies mainly rely on artificial rules or static relationship templates for semantic mapping of parameters and objectives and conflict explanation, and lack an extensible conflict explanation framework based on data-driven and industrial knowledge fusion. For cases such as expansion of production line scale, online of new equipment, or coupling of complex processes, the parameter conflict model is difficult to adaptively upgrade, resulting in limitations in model reliability and transparency; Furthermore, due to limitations in the expressive power of existing models, conflict contribution, impact paths, and actionable recommendations often cannot be systematically generated. Many production systems still rely on experience-based adjustments and manual intervention for optimization decisions, lacking structured conflict diagnostic reports and strategy recommendations based on quantitative indicators. This results in industry pain points such as "opaque optimization mechanisms, difficulty in tracing the source of decisions, and difficulty in implementing intervention recommendations." Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a method for intelligent optimization of packaging parameters in a packaging production line.

[0004] The technical solution of this invention is implemented as follows: A method for intelligent optimization of packaging parameters in a packaging production line, comprising: S1: Collect multi-source parameter data during the operation of the packaging production line. The multi-source parameter data includes sealing temperature, operating speed, tension control, energy consumption indicators and ambient temperature and humidity. Divide the data into time series according to three time granularities: minute level, hour level and batch level to form a multi-scale raw dataset. S2: Standardize and denoise the original multi-scale dataset, and use the sliding window method to extract the statistical and fluctuation features at each time granularity to obtain the basic change patterns of each parameter at different time scales. S3: Construct a multi-scale feature encoder, decouple the basic change patterns based on a self-supervised learning model, and extract the independent change patterns of parameters on the time scale, including instantaneous fluctuation effects and steady-state offset effects, to generate multi-scale feature vectors. S4: Input the multi-scale feature vector into the functional semantic mapping module, and combine it with the preset process target label to map each parameter to its corresponding process target, including quality stability, energy consumption efficiency and equipment life, so as to generate a parameter-target correlation matrix; S5: Based on the parameter-target correlation matrix and multi-scale feature vector, execute a dual-scale conflict detection algorithm, in which local conflicts caused by instantaneous disturbances are identified on the time scale, and global conflicts caused by essential contradictions between targets are identified on the functional scale, so as to generate conflict type labels and influence paths; S6: Based on the conflict type label, impact path, and parameter-target correlation matrix, calculate the contribution weight ranking of each conflict source and generate a structured conflict report, which includes conflict type, impact path, weight ranking, and actionable recommendations. S7: Based on the structured conflict report, determine whether the preset optimization intervention threshold is met. If it is met, output parameter dynamic response threshold adjustment suggestions or target priority reconstruction strategies for manual intervention or automatic strategy update module to execute. S8: After the strategy is executed, continuously monitor the operating status of the packaging production line, collect feedback data and compare it with historical data, and update the parameter configuration of the multi-scale feature encoder and functional semantic mapping module to achieve dynamic optimization and adaptive improvement of conflict modeling capabilities.

[0005] The present invention provides a method for intelligent optimization of packaging parameters in a packaging production line, which has the following beneficial effects: (1) This invention decouples parameter changes in the packaging production line into two major feature components: instantaneous disturbance and steady-state offset, through multiple time granularities such as minutes, hours, and batches. Combined with semantic mapping of functional objectives (such as quality, energy consumption, and lifespan), it directly provides the cause and action chain of the conflict, realizing a structured explanation and path visualization analysis of parameter conflicts. Compared with existing black-box algorithms, the system transparency and user trust are significantly improved, providing accurate and easy-to-understand factual basis for subsequent manual optimization and intelligent decision-making; (2) Through self-supervised multi-scale feature encoding and wavelet packet decomposition, this invention can accurately distinguish between local time-varying anomalies (such as short-term instability of a single sealing operation) and long-period steady-state shifts (such as continuous high energy consumption trends), and uses attention mechanisms, sliding window difference algorithms, and other algorithms to achieve high-precision capture of data mutations and trend differentiation. At the functional scale, the causal and priority evaluation of process objectives is introduced, enabling quantitative identification of the essential conflicts between different optimization objectives. Simultaneously, an efficient inference model ensures online real-time conflict alarms and responses, solving the problem of the lag in traditional algorithms that can only summarize data after the fact. (3) This invention uses parameter-target semantic strength and influence weight as a link to couple multi-scale features with conflict type, path hierarchy and target priority. Through matrix multiplication and influence distribution normalization, it achieves the ranking of the contribution of each conflict under all targets and the output of the influence path topology. It can not only indicate "what conflict was generated", but also clearly indicate "who is the main cause, which target was affected, and how to adjust". This capability directly overcomes the limitation of existing technologies that can only output the overall weight and cannot refine the accountability and optimization objects, greatly enhancing the application value of the system for actual production management. Attached Figure Description

[0006] Fig. 1 This is a flowchart of a method for intelligent optimization of packaging parameters in a packaging production line according to the present invention; Fig. 2 This is a sub-flowchart of a method for intelligent optimization of packaging parameters in a packaging production line according to the present invention; Fig. 3 This is another sub-flowchart of the intelligent optimization method for packaging parameters in a packaging production line according to the present invention. Detailed Implementation

[0007] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0008] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0009] When used herein, the singular forms of “a,” “an,” and “the” may also include the plural forms unless the context clearly indicates otherwise. It should also be understood that terms such as “comprising / including” or “having” specify the presence of the stated features, integrals, steps, operations, components, parts, or combinations thereof, but do not preclude the possibility of the presence or addition of one or more other features, integrals, steps, operations, components, parts, or combinations thereof. Meanwhile, the term “and / or” as used in this specification includes any and all combinations of the associated listed items.

[0010] Please see Figs. 1-3 As shown, a method for intelligent optimization of packaging parameters in a packaging production line includes: S1: Collect multi-source parameter data during the operation of the packaging production line. The multi-source parameter data includes sealing temperature, operating speed, tension control, energy consumption indicators and ambient temperature and humidity. Divide the data into time series according to three time granularities: minute level, hour level and batch level to form a multi-scale raw dataset. S2: Standardize and denoise the original multi-scale dataset, and use the sliding window method to extract the statistical and fluctuation features at each time granularity to obtain the basic change patterns of each parameter at different time scales. S3: Construct a multi-scale feature encoder, decouple the basic change patterns based on a self-supervised learning model, and extract the independent change patterns of parameters on the time scale, including instantaneous fluctuation effects and steady-state offset effects, to generate multi-scale feature vectors. S4: Input the multi-scale feature vector into the functional semantic mapping module, and combine it with the preset process target label to map each parameter to its corresponding process target, including quality stability, energy consumption efficiency and equipment life, so as to generate a parameter-target correlation matrix; S5: Based on the parameter-target correlation matrix and multi-scale feature vector, execute a dual-scale conflict detection algorithm, in which local conflicts caused by instantaneous disturbances are identified on the time scale, and global conflicts caused by essential contradictions between targets are identified on the functional scale, so as to generate conflict type labels and influence paths; S6: Based on the conflict type label, impact path, and parameter-target correlation matrix, calculate the contribution weight ranking of each conflict source and generate a structured conflict report, which includes conflict type, impact path, weight ranking, and actionable recommendations. S7: Based on the structured conflict report, determine whether the preset optimization intervention threshold is met. If it is met, output parameter dynamic response threshold adjustment suggestions or target priority reconstruction strategies for manual intervention or automatic strategy update module to execute. S8: After the strategy is executed, continuously monitor the operating status of the packaging production line, collect feedback data and compare it with historical data, and update the parameter configuration of the multi-scale feature encoder and functional semantic mapping module to achieve dynamic optimization and adaptive improvement of conflict modeling capabilities.

[0011] Step S1: Collect multi-source parameter data during the operation of the packaging production line. This multi-source parameter data includes sealing temperature, operating speed, tension control, energy consumption indicators, and ambient temperature and humidity. The data is then divided into time series at three time granularities: minute, hour, and batch, to form a multi-scale raw dataset. Specifically, this includes: S1.1: Based on the industrial Internet of Things platform, acquire multi-source parameter data during the operation of the packaging production line. The multi-source parameter data includes sealing temperature, operating speed, tension control, energy consumption index and ambient temperature and humidity, to obtain the original time-series data stream. Based on the industrial IoT platform, the data acquisition interface (parameter: sampling frequency 1Hz~10Hz) deployed in the sealing temperature monitoring unit, running speed sensor, tension control module, energy consumption metering device and environmental temperature and humidity detector is called to realize the real-time acquisition of multi-source parameters during the operation of the packaging production line. Furthermore, the raw signals generated by the aforementioned sensors are transmitted to a centralized data access gateway through the standardized MQTT industrial messaging protocol (parameters: QoS level=2, topic structure follows the process section / equipment ID / parameter type format), thereby realizing unified data access and secure transmission across devices and plant areas; Furthermore, a timestamp synchronization mechanism (parameter: NTP network clock synchronization accuracy ≤ 10ms) is adopted to add high-precision unified time tags to the multi-source parameter data accessed through the gateway, and generate original data records with <device ID, parameter type, timestamp, value> as the four-tuple to ensure the comparability of time series and subsequent alignment processing; Furthermore, a message queue buffering mechanism is used (parameters: memory queue capacity ≥ 1GB, packet loss rate ≤ 10%). Asynchronous caching of multi-source data from different acquisition cycles enables smoothing of the data stream and flow control management, reducing acquisition gaps caused by transmission jitter; Furthermore, a CRC32-based integrity verification algorithm is used to perform verification calculations on the original data packets in the buffer queue, thereby realizing the detection of bit errors in the data during the acquisition and transmission process, marking verification failure records, and sending them to the abnormal data annotation list; Through the above-mentioned industrial IoT data acquisition and unified access processing method, the heterogeneous signals from multiple sensors in the previous step are transformed into raw time-series data streams with unified structure, comparable time, and complete verification, so as to achieve the expected technical effects of subsequent time alignment, granular slicing and feature extraction. For example, on a beverage packaging production line with a rated capacity of 120 units per minute, the following sensors are configured: a TCS-800 thermocouple sealing temperature sensor (sampling frequency 5Hz, accuracy ±0.5℃), an encoder-type running speed sensor (sampling frequency 10Hz, resolution 0.01 m / s), a strain gauge tension sensor (sampling frequency 2Hz, range 0-500N), a power meter-type energy consumption monitoring device (sampling frequency 1Hz, accuracy ±0.5%), and a DHT-22 temperature and humidity sensor (sampling frequency 0.5Hz, temperature accuracy ±0.3℃, humidity accuracy ±2%RH). All sensors are connected via an industrial IoT gateway supporting the MQTT protocol, with a QoS level of 2 and a topic format of / line01 / deviceID / paramType. The gateway synchronizes with an atomic clock source via NTP service, achieving a time synchronization accuracy of 5ms. Data enters a 2GB memory buffer queue, with a packet loss rate of less than 10%. The data was checked using CRC32, resulting in a zero checksum error rate. In the final generated raw time-series data stream, each record contains a device ID, such as the sealing temperature sensor ID=TEMP001, parameter type=TempSeal, timestamp=2024-06-08 14:35:21.005, and value=185.4. This data stream supports timestamp alignment and unified sampling rate processing in S1.2, ensuring data continuity and consistency across multiple scales (minute, hour, and batch levels), effectively improving the quality of input data in the conflict modeling and analysis phase. S1.2: Perform timestamp alignment and sampling rate unification processing on the original time-series data stream, and use linear interpolation and sliding window mean filtering algorithms to repair missing values ​​and outliers in order to generate a time-aligned structured original dataset; Based on the original time-series data stream collected by the industrial IoT platform and verified by CRC32, a unified time reference framework is used to perform timestamp alignment processing on all data records (parameter: NTP synchronization accuracy ≤ 10ms), so as to achieve accurate mapping of cross-device data points on a unified time axis. Furthermore, a sampling rate resampling algorithm (parameter: target sampling rate = ...) is used. =5Hz) The sampling frequency of different source parameters is normalized and adjusted. Using a multi-channel synchronous resampling scheme, the high-frequency sampled data is interpolated into the low-frequency synchronization within the period, and the low-frequency sampled data is undersampled and interpolated to make up the gap, so as to align the full parameter sequence at the target sampling rate. Furthermore, for short-term missing points in the time-aligned parameter sequences, a linear interpolation method is used to smoothly fill single points or a small number of continuous gaps, calculated as follows: in , The values ​​are those of adjacent known sampling points. , For adjacent known sampling times, For the missing point time; Furthermore, to address abnormal peak values ​​caused by equipment jitter or sensor transient distortion during continuous data acquisition, a sliding window mean filtering algorithm (parameters: window length w = 5 points, sliding step s = 1 point) is applied to calculate the mean within the current window. It replaces the outlier at the center point with it, thereby smoothing out spike interference and preserving the trend component; Furthermore, a consistency check is performed on the interpolated and filtered data sequence, and the absolute value of the difference between adjacent samples is calculated. The mean and standard deviation exceed Points with ±3σ are marked as potential residual anomalies for subsequent feature extraction and removal. By using timestamp alignment, sampling rate unification, linear interpolation completion, and moving average denoising, the original time series data stream from the previous step is transformed into a structured original dataset with consistent time reference, unified sampling frequency, and missing and anomaly repair, achieving the high integrity and comparability technical effects required for multi-scale time series slicing. For example, in the data processing of a beverage packaging production line with a rated capacity of 120 units per minute, the sampling frequency for the sealing temperature sensor is 5Hz, the operating speed sensor is 10Hz, the tension sensor is 2Hz, the energy consumption monitoring is 1Hz, and the ambient temperature and humidity is 0.5Hz. The target sampling rate is set to 5Hz. Sensors with increased sampling rates (energy consumption and environmental parameters) use linear interpolation to fill in the gaps, while those with decreased sampling rates (speed sensor) use average across different sampling points. For example, if the sealing temperature is missing at 2024-06-08 14:35:25.200, this point is filled by the previous available point. and the next available point Based on the time ratio, it is calculated as follows: In the moving average filtering (w=5), a peak value of 5.50 m / s was detected in the running speed sequence, which was >3σ higher than the surrounding mean of 5.20 m / s. This peak value was replaced with the window mean of 5.21 m / s. The processed 5Hz structured raw dataset for each channel showed no gaps or peaks within a continuous 1-hour data batch, and the standard deviation decreased by 12%, providing high-quality input for minute-level, hour-level, and batch-level statistical analysis in S1.3 to S1.5. S1.3: The structured original dataset is sliced ​​into time windows based on minute-level time granularity to extract the instantaneous values ​​and statistical features of each parameter at the minute-level granularity, so as to generate a minute-level time series data subset; S1.4: Perform sliding window aggregation calculation on the structured original dataset based on hourly time granularity, and use mean, variance and range statistical methods to extract the trend features of each parameter at the hourly granularity to generate an hourly time series data subset; S1.5: Divide the structured raw dataset into batches based on batch-level production cycle identifiers, and extract the overall distribution characteristics of each parameter at the batch-level granularity by combining the start and end timestamps of the batches, so as to generate a batch-level time series data subset. S1.6: The minute-level, hour-level, and batch-level time series data subsets are fused and formatted to generate a multi-scale original dataset containing multiple time granularity labels for use in subsequent standardization and feature extraction steps.

[0012] Step S2: The multi-scale original dataset is standardized and denoised preprocessed, and the sliding window method is used to extract statistical and fluctuation features at each time granularity to obtain the basic change patterns of each parameter at different time scales. Specifically, this includes: S2.1: Missing values ​​are detected and filled in the parameters such as sealing temperature, running speed, tension control, energy consumption index and ambient temperature and humidity in the multi-scale original dataset. The missing data is filled in by linear interpolation method to generate a complete time series dataset, so as to ensure the continuity and consistency of subsequent feature extraction. In the multi-scale raw dataset generated based on step S1.6, time series parameters such as sealing temperature, running speed, tension control, energy consumption index and ambient temperature and humidity are selected as processing objects, with the goal of generating a complete time series dataset that is free of missing data and continuous and consistent. A missing value detection algorithm (parameters: detection window length = 10 points, threshold ratio = 1) is used to scan each parameter sequence point by point to identify the missing value location and duration, and output a missing value index list. Furthermore, numerical completion is performed on the detected missing regions using a linear interpolation method (parameter: interpolation order = 1). The calculation method is as follows: in and The values ​​are the values ​​of adjacent known points. and The interpolation result is obtained by proportional weighting for the corresponding timestamp; Furthermore, for continuous missing segments spanning multiple sampling points, piecewise linear interpolation is used to reconstruct the values ​​of each intermediate point, ensuring a smooth transition of long gaps and maintaining the continuity of parameter change trends; Furthermore, boundary consistency correction is performed on the interpolated data sequence. For cases where the beginning or end of the sequence is missing, the nearest neighbor extrapolation method is used to fill the outermost sampling point with the nearest valid value to avoid NaN or null values ​​affecting subsequent feature calculations. Furthermore, data integrity verification is performed on the padded sequence, and the missing rate metric is calculated: in This is the total length of the original sequence. The number of valid points after filling, when When the value approaches 0, it is marked as a complete sequence; By combining missing value detection with linear interpolation completion and boundary extrapolation, the time series with gaps in the multi-scale original dataset of the previous stage are transformed into a consistent dataset with full point filling and continuous time, so as to achieve the high coherence and distortion-free input effect required for subsequent standardization and feature extraction. For example, in a beverage packaging production line with a rated capacity of 120 units per minute, the sealing temperature sensor (sampling frequency 5Hz) shows two scattered missing data points (located at 14:35:25.200 and 14:35:40.600 on 2024-06-08) and one continuous gap (from 14:35:50.000 to 14:35:50.400, over two cycles). The missing value detection window length is set to 10 points, and the threshold ratio is 1. After detecting the gap index, single-point missing values ​​are processed through adjacent points... and Calculate the interpolation value Another point is... and Interpolation is Continuous gaps are identified through piecewise interpolation, utilizing the endpoints. and Generate the two middle points and The sequence is complete from beginning to end, and no extrapolation was called. The imputation rate is 100%, and the missing rate is... This ensures that the complete time series can be normalized to Z-score in S2.2, improving the stability and comparability of subsequent multi-scale feature extraction; S2.2: Based on the complete time series dataset, standardization processing is performed on each parameter. The Z-score normalization method is used to center and scale the data to eliminate the differences in the units of different parameters, generate a standardized parameter sequence, and improve the stability and comparability of the feature extraction process. S2.3: Apply wavelet threshold denoising algorithm to denoise the standardized parameter sequence, and suppress high-frequency noise based on wavelet transform coefficients to obtain the denoised time series signal, thereby improving the signal-to-noise ratio and interpretability of parameter change patterns. S2.4: Based on the denoised time series signal, the sliding window method is used to set the window length and step size for three time granularities: minute, hour, and batch. The time window is divided in minutes, and the sliding step size is 30 seconds. Local statistical features at each time granularity are extracted, including mean, variance, kurtosis and skewness, to generate statistical feature vectors at multiple time granularities. S2.5: Further extract fluctuation features from the statistical feature vectors at each time granularity, calculate the rate of change and cumulative change of feature values ​​between adjacent windows, in order to capture the dynamic change trend of parameters at different time scales, and generate a basic change pattern feature set as input data for subsequent multi-scale feature encoders.

[0013] Step S3: Construct a multi-scale feature encoder, decouple the basic change patterns based on a self-supervised learning model, and extract the independent change patterns of parameters on the time scale, including instantaneous fluctuation effects and steady-state shift effects, to generate multi-scale feature vectors. For example... Fig. 2 As shown, it specifically includes: S3.1: Based on the aforementioned basic change pattern, a self-supervised learning framework is constructed, and a temporal comparison learning strategy is adopted to extract unlabeled features from parameter change patterns at multiple temporal granularities, so as to construct a potential feature space for parameters at different time scales. S3.2: Perform scale separation operation on the potential feature space, use multi-scale wavelet packet transform to decompose the feature vector in the frequency domain, divide the parameter change mode into high-frequency instantaneous fluctuation component and low-frequency steady-state offset component, so as to achieve feature decoupling in the time scale dimension. For the potential feature space, parameter change feature vectors composed of multiple time granularities such as minute, hour, and batch are selected as input objects, with the goal of achieving feature decoupling in the time scale dimension; A multi-scale wavelet packet transform algorithm (parameters: number of decomposition levels = 3, wavelet basis function = Daubechies-4) is used to decompose the input feature vector into a full-band decomposition, thereby achieving fine division of energy distribution in different frequency bands. Furthermore, by calculating the frequency range of the coefficients of each wavelet packet node, frequencies higher than a set high-frequency threshold are selected. The node coefficients are classified as high-frequency transient fluctuation components, with frequencies below a set low-frequency threshold. The node coefficients are classified as low-frequency steady-state offset components; Furthermore, energy normalization is applied to standardize the coefficient energy of each frequency band, using the following formula: in This represents the energy of the node. This represents the total energy across the entire frequency band, thus providing the relative energy distribution characteristics of high and low frequency components; Furthermore, by reconstructing the formula High frequency coefficients With the corresponding wavelet basis Reconstructing into a transient fluctuation signal, reducing the low-frequency coefficients With the corresponding wavelet basis Reconstructed into a steady-state offset signal; Furthermore, by utilizing the cross-correlation coefficients between high and low frequency signals... Assess the independence of the two types of components to ensure the effectiveness of feature decoupling; By using the above-mentioned multi-scale wavelet packet decomposition and frequency band classification methods, the latent feature space is effectively separated into instantaneous fluctuation components and steady-state offset components, realizing feature decoupling in the time scale dimension, and laying a high-resolution frequency domain input foundation for subsequent instantaneous disturbance feature and steady-state offset feature encoding. For example, on a food packaging production line with a rated capacity of 300 pieces per minute, the minute-level sealing temperature sequence obtains a 256-dimensional latent feature vector in step S3.1. Using three-layer wavelet packet decomposition and the Daubechies-4 wavelet basis, the entire frequency band of the sequence is divided into 8 sub-bands, and a high-frequency threshold is set. =2.5Hz, low frequency threshold =0.4Hz, the energy of the sub-band coefficient above the high-frequency threshold accounts for 26%, which is denoted as the instantaneous fluctuation component, and the energy below the low-frequency threshold accounts for 41%, which is denoted as the steady-state offset component. The cross-correlation coefficient between the two is... =0.08, indicating high independence. The reconstructed instantaneous fluctuation signal improves the sensitivity of detecting local temperature control anomalies by 15%, and the steady-state offset signal improves the stability of monitoring the thermal inertia trend of the equipment by 20%, providing high-quality feature input for distinguishing between time-varying disturbances and functional target conflicts in subsequent conflict detection; S3.3: Based on the instantaneous fluctuation component and the steady-state offset component, a scale feature encoder is designed, and a long short-term memory network is used to perform local temporal modeling of the high-frequency fluctuation component to generate instantaneous disturbance feature vectors; Based on the separation results of instantaneous fluctuation components and steady-state offset components, high-frequency instantaneous fluctuation components are selected as the main input object of the scale feature encoder; A Long Short-Term Memory (LSTM) network (parameters: number of hidden layer units = 128, time step = 20, dropout rate = 0.2) is used to perform local time series modeling on high-frequency instantaneous fluctuation signals, capturing their rate abrupt changes and amplitude transitions between adjacent time segments. Furthermore, an optimization strategy is implemented through gradient pruning (parameter: pruning threshold). To suppress the gradient explosion problem during backpropagation of long sequences, ensure that the network maintains convergence stability when modeling high-frequency perturbations, and output a preliminary instantaneous perturbation temporal feature matrix; Furthermore, a multi-head attention mechanism (parameters: number of attention heads = 4, key / value dimension = 32) is used to redistribute the weights of the transient perturbation feature matrix, thereby increasing the contribution weight of key perturbation segments in the feature representation and generating a weighted transient perturbation feature vector. Furthermore, the weighted instantaneous perturbation feature vector is standardized by using a batch normalization layer to reduce the training instability caused by the inconsistency of feature value ranges in different batches, and to ensure that the encoding results are aligned in subsequent feature fusion. By jointly encoding with a long short-term memory network and an attention mechanism, high-frequency instantaneous fluctuation components are transformed into instantaneous perturbation feature vectors with fixed dimensions and rich local perturbation patterns, thereby achieving high-precision representation of rapidly changing processes. For example, in a daily chemical product filling and packaging production line with a rated capacity of 150 units per minute, the length of the minute-level sealing temperature high-frequency instantaneous fluctuation signal is 1200 points. The LSTM network time step size is set to 20, and the corresponding input sequence is divided into 60 time step windows. In the network forward propagation, the input vector dimension of each time step is 16, the number of hidden layer units is 128, the dropout rate is 0.2, and the gradient pruning threshold is used. This avoids the gradient norm exceeding 1.2 in the early stages of training. The multi-head attention mechanism calculates a weighted score matrix with four heads, and the key / value vector has a dimension of 32. After Softmax normalization, it significantly improves the perturbation weights between time steps 15 and 18, increasing the overall attention concentration ratio from 0.42 to 0.67. The output after batch normalization is an instantaneous perturbation feature vector of length 128. In the subsequent conflict detection task, this vector improved the accuracy of identifying quality target conflicts caused by closure anomalies by 13.5%. S3.4: Perform a trend extraction operation on the steady-state offset component, and use a method combining sliding window averaging and polynomial fitting to extract the long-term evolution trend of the parameters to generate a steady-state offset feature vector. S3.5: The instantaneous disturbance feature vector and the steady-state offset feature vector are concatenated and fused, and a fully connected neural network is used to perform nonlinear mapping to generate a unified multi-scale feature vector, which is used as the input of the subsequent functional semantic mapping module.

[0014] Step S4: Input the multi-scale feature vector into the functional semantic mapping module, and combine it with the preset process target label to map each parameter to its corresponding process target, including quality stability, energy efficiency, and equipment life, to generate a parameter-target correlation matrix. For example... Fig. 3 As shown, it specifically includes: S4.1: Perform target label alignment operation on each dimension of the parameter feature in the multi-scale feature vector. Based on the preset process target classification system, perform preliminary semantic matching between the parameter features and three types of process targets: quality stability, energy efficiency and equipment life, to generate a preliminary parameter-target mapping table. Based on multi-scale feature vector input, a target label alignment algorithm (parameters: number of label categories = 3, category set = {quality stability, energy efficiency, equipment life}) is adopted to realize the semantic correspondence between the features of each parameter and the process target classification system. Furthermore, by using the feature attribute extraction method (parameters: statistical index set = {mean, variance, kurtosis, skewness}), the core numerical features representing the change pattern of physical quantities in the multi-scale feature vector are extracted, and this is used as the basic feature set for semantic matching. Furthermore, the Euclidean distance matching algorithm is used to calculate the similarity between the basic feature set and the typical feature templates of each process target. The template library is derived from historical operation big data statistics and target feature patterns manually annotated by experts. Furthermore, similarity normalization unifies the similarity values ​​of different parameter-target combinations to the [0,1] range, ensuring the comparability of cross-parameter and cross-target matching strengths; Furthermore, the highest normalized similarity target label corresponding to each parameter feature is selected using the maximum matching rule to generate a set of parameter-target matching pairs; By constructing a preliminary parameter-target mapping table using the above matching pair set, a one-to-one correspondence between parameter features and process targets is achieved, providing a structured mapping foundation for subsequent semantic enhancement and conflict detection; For example, in a beverage packaging production line with a rated capacity of 180 units per minute, the multi-scale feature vector contains the mean of instantaneous perturbations in sealing temperature over minutes. Hourly energy consumption bias Batch-level tension offset kurtosis Features such as... In the constructed process target template, the mean range of the quality stability target is... Furthermore, the kurtosis is close to 2, and the skewness of the energy efficiency target is concentrated in... Within this range, the tension kurtosis range corresponding to the equipment lifespan target is within The differences between the sealing temperature characteristic and the mean value of the quality template, the energy consumption skewness and the energy consumption template value, and the tension kurtosis and the lifetime template value were calculated using the Euclidean distance formula, respectively, to obtain the original similarity values ​​as follows: , , After normalization, they are respectively , , The final matching rules selected from sealing temperature to quality stability, energy consumption to energy efficiency, and tension to equipment life. The generated preliminary mapping table accurately identified the main process target attributes of each parameter, providing a reliable basic input for the subsequent steps of introducing domain knowledge graph enhancement. S4.2: Based on the process knowledge graph, perform semantic enhancement processing on the parameter-target preliminary mapping table, and use the parameter-target causal rules defined by domain experts to correct and expand the mapping relationship to generate an enhanced parameter-target semantic mapping table; Based on the parameter-target preliminary mapping table input, a process knowledge graph reasoning algorithm (parameters: number of knowledge graph nodes > 5000, number of relation types = 12) is used to retrieve and deduce the implicit semantic relationship between parameter features and process targets. Furthermore, by using the causal rule matching method (parameters: rule base size = 320 rules, rule form = premise-result tuple), the parameter-target causal chain defined by domain experts is used to verify the logical consistency of the matching pairs in the preliminary mapping table, and missing or indirect parameter-target associations are derived according to the rules to expand the mapping coverage. Furthermore, a semantic conflict detection algorithm is used to perform conflict analysis on the mapping relationship after causal rule derivation, and the rule consistency coefficient of each parameter-target pair is calculated. The formula is as follows: in The number of matches to conform to the causal rule. This represents the total number of matches, used to quantify the reliability of the mapping rules. Elimination. Weak reliability mapping pairs below 0.6; Furthermore, a multi-hop inference mechanism (parameters: maximum number of hops = 3, path weight decay coefficient = 0.85) is employed to perform cross-node association search in the process knowledge graph, capturing parameters and targets that are not directly connected but have high semantic relevance. This is achieved through a path relevance scoring formula: in For path weights, To calculate path similarity, extract additional semantic completion associations; Furthermore, the results of the initial mapping, causal inference completion, and multi-hop reasoning completion are weighted and synthesized by a weighted fusion algorithm (parameters: knowledge graph weight ratio 0.6, causal rule weight ratio 0.4) to form an enhanced parameter-target semantic mapping table. By using the above-mentioned knowledge graph and causal rule fusion reasoning method, the initial mapping relationship is optimized into an enhanced mapping matrix that has both process logic consistency and semantic comprehensive coverage, so as to realize the semantic association input from high-quality parameters to process objectives before multi-objective conflict modeling; For example, on a beverage filling line with a rated capacity of 220 units per minute, the initial mapping table recorded three direct matches: sealing temperature → quality stability (similarity 0.968), operating speed → energy efficiency (similarity 0.915), and tension control → equipment lifespan (similarity 1.0). The process knowledge graph has 7480 nodes, including equipment parameter nodes, process target nodes, and control strategy nodes, covering 12 types of relationships such as influence, constraint, and dependency. A rule defined by domain experts, such as "If parameter X has a positive impact on target Y, and target Y and target Z are mutually exclusive, then parameter X has a negative impact on target Z," is numbered R102 in the rule base, corresponding to a premise pattern matching rate of 92%. For the sealing temperature → quality stability match, the causal rule expands to show a weak negative association between sealing temperature and energy efficiency (association strength 0.42), and completes the indirect positive association between sealing temperature and equipment lifespan (association strength 0.65). Conflict detection calculation. Sealing temperature → Energy efficiency =0.58, rejected; Sealing temperature → Equipment lifespan =0.81 is retained. Multi-hop inference (maximum number of hops = 3, attenuation coefficient 0.85) revealed a highly correlated path between operating speed and equipment lifespan nodes via equipment load and maintenance cycle nodes (path weight 0.72, similarity 0.88, path score 0.634), thus completing the correlation between operating speed and equipment lifespan. Weighted fusion calculation of the final correlation weight between sealing temperature and equipment lifespan was then performed. Operating speed → Equipment lifespan weight The enhanced semantic mapping table generated after processing effectively covers key indirect influence paths, providing a more comprehensive and reliable parameter-target association input for the subsequent semantic similarity quantification calculation in S4.3; S4.3: For each parameter-target pair in the enhanced parameter-target semantic mapping table, the semantic similarity calculation model is used to calculate its semantic association strength. The semantic similarity calculation model is trained based on the terminology database of the industrial control field to generate the parameter-target association strength matrix. Based on the input of the enhanced parameter-target semantic mapping table, a semantic similarity calculation model (parameter: vector dimension = 300, training corpus covers equipment, process and operation and maintenance terms in the field of industrial control, model type = Word2Vec Skip-Gram) is used to numerically represent each parameter-target pair, so as to realize the vectorized encoding of the semantic features of parameters and targets; Furthermore, using the vectorized encoding results, the cosine similarity calculation method is employed to quantify the semantic closeness between the parameters and the target. Furthermore, the cosine similarity results are normalized using the min-max normalization formula. To eliminate the differences in the original similarity distribution between different parameter-target pairs, a standardized association strength value between 0 and 1 is obtained, where and These are the minimum and maximum values ​​among all calculation results in the current batch, respectively. Furthermore, to improve the accuracy of representing long-tail terms and compound words unique to the field of industrial control, a sub-word embedding enhancement strategy (parameter: n-gram range = 3~6) is adopted. Based on the original word vectors, a weighted superposition of sub-word vectors is introduced to optimize the semantic similarity calculation of rare parameter nouns or target descriptions. Furthermore, all normalized semantic similarity values ​​are filled into the corresponding cells of the association matrix according to the parameter and target classification index to generate the parameter-target association strength matrix. The rows of this matrix are the parameter set, and the columns are the target set. The matrix elements take values ​​in the range [0,1]. Through the above semantic similarity calculation and enhancement processing, the enhanced semantic mapping table is transformed into a quantifiable and comparable parameter-target association strength matrix, achieving a high-precision measurement effect of semantic coupling between different process elements; For example, on a pharmaceutical packaging production line with a rated capacity of 250 pieces per minute, the enhanced semantic mapping table includes mapping relationships such as "sealing pressure" → "quality stability," "conveyor belt speed" → "energy efficiency," and "cooling airflow temperature" → "equipment lifespan." A Skip-Gram word vector model (300 dimensions) trained on 7.8 million words from the industrial control domain is used to encode each parameter and target after word segmentation. For example, the vector for "sealing pressure" is... The corresponding "quality stability" vector is Substituting the values ​​into the cosine similarity formula yields S = 0.945. After normalization (maximum value 0.97, minimum value 0.61 for the batch), the result is... =0.902. For sub-word granularity processing, terms such as "cooling airflow temperature" are generated into n-gram sub-word fragments of length 4 to 6, which are then added to the original vector representation, increasing its similarity with the target "equipment lifespan" from 0.71 to 0.79, and after normalization, to 0.75. The results are then filled into matrix cells according to the parameter × target axis. For example, matrix (sealing pressure, quality stability) = 0.902, (cooling airflow temperature, equipment lifespan) = 0.75. The resulting association strength matrix weakens parameter-target pairs with coupling less than 0.4 in subsequent conflict detection, achieving accurate conflict source identification and optimization strategy formulation. S4.4: Based on the parameter-target correlation strength matrix, a normalization algorithm is used to standardize the correlation strength of each parameter under different process targets to generate a standardized parameter-target correlation matrix; S4.5: Perform sparsification on the standardized parameter-target association matrix, set a semantic association threshold, and reset the parameter-target association weights below the threshold to zero to generate a sparsified parameter-target association matrix, which serves as the input for the structured conflict detection module.

[0015] Step S5: Based on the parameter-target correlation matrix and multi-scale feature vectors, a dual-scale conflict detection algorithm is executed. This algorithm identifies local conflicts caused by instantaneous disturbances at the time scale and global conflicts caused by fundamental contradictions between targets at the functional scale, generating conflict type labels and impact paths. Specifically, this includes: S5.1: Perform target dimension projection transformation on the parameter-target correlation matrix, group the multi-scale feature vectors based on the target semantic label to separate the parameter change response sequence under each process target, and obtain the feature projection vector under the target dimension for subsequent functional scale conflict modeling. S5.2: The sliding window difference analysis method is used to detect local mutations in the feature vectors under the time scale, calculate the feature difference degree between adjacent time windows, identify abnormal fluctuations in parameters caused by instantaneous disturbances, and generate a time disturbance intensity index for local conflict identification. In time-scale conflict detection, the input data is the feature projection vector of the target dimension output by step S5.1, which includes multi-scale parameter change response sequences at the minute, hour, and batch levels. The sliding window difference analysis method (parameters: window length = 3 sampling points, sliding step size = 1 sampling point) is used to capture and quantify local changes in the time series. Furthermore, the formula for calculating the difference in mean values ​​between adjacent time windows is as follows: To extract the mean difference between consecutive windows, where... The feature mean of the i-th window; Furthermore, the variance difference between adjacent time windows is calculated using the following formula: To measure the degree of fluctuation, where Let be the standard deviation of the features in the i-th window; Furthermore, the comprehensive difference index calculation formula is adopted: Achieve comprehensive quantification of the amplitude and volatility changes of instantaneous disturbances; Furthermore, a threshold determination method (parameter: threshold ratio coefficient = 1.5 times the mean difference) is used to identify time periods where the overall difference exceeds the normal fluctuation range, in order to mark instantaneous disturbance events; Furthermore, the disturbance intensity index is calculated for all transient disturbance events: in To quantify the risk of local conflicts by measuring the number of disturbance events; By using differential analysis and threshold detection, the temporal variation pattern of the feature vector is transformed into a temporal perturbation intensity index, thus realizing the quantitative input required for local conflict identification. For example, on a food packaging production line with a rated capacity of 200 pieces per minute, the minute-level sealing temperature characteristic sequence, after being processed by S5.1, has a sliding window length of 3 minutes and a step size of 1 minute. The average value of adjacent windows is calculated to obtain... They are respectively ℃ ℃ ℃, variance They are respectively , , The overall difference is calculated as follows: , , When the threshold ratio is set to 1.5 and the mean difference between normal and normal values ​​is... ℃, then the threshold is ℃, detected the comprehensive difference in the second window Events exceeding the threshold are marked as transient disturbances. The number of disturbances within one hour is counted. =3, Disturbance Intensity Index = = This provides a quantitative time perturbation input for subsequent multi-scale conflict detection, enabling timely early warning of local conflicts caused by abnormal fluctuations in sealing temperature; S5.3: Based on the priority constraint relationship between process objectives and the objective achievement function, a multi-objective conflict scoring model under the functional scale is constructed. The conflict score is calculated on the feature projection vector under the objective dimension to quantify the contradiction intensity between each objective and generate a functional conflict scoring matrix. In the functional scale conflict detection step, the input data are the target dimension feature projection vector obtained in step S5.1, and the sparsified parameter-target correlation matrix generated by S4.5. The conflict score is calculated by combining the priority constraint relationship between process targets and the target achievement function. The priority constraint weight allocation method is adopted (parameters: priority weight range [0,1], weight normalization constraint). This allows for the quantification of the importance of each process objective (quality stability, energy efficiency, equipment lifespan), providing prior weighting coefficients for conflict scoring. Furthermore, the achievement level of each process objective under the current feature projection is calculated using a target achievement degree function, which is expressed in functional form. ,in For the current observation value, Set values ​​for the target to achieve a standardized representation of the target's completion rate; Furthermore, a multi-target conflict relative deviation calculation method is adopted (parameter: conflict sensitivity coefficient α∈[0,2]), using the formula... To quantify the relative conflict intensity between any two targets i and j, where Elements of the conflict scoring matrix; Furthermore, through matrix filling operations, the conflict intensity values ​​of all target pairs are filled into the functional conflict scoring matrix according to the row and column correspondence rules, and diagonal element nulling is performed to ensure that the matrix only reflects cross-conflict relationships; Furthermore, to suppress the non-steady-state impact of data noise on conflict scoring, exponential smoothing (parameter: smoothing coefficient λ=0.3) is used to smooth the scoring matrix over time. The update formula is as follows: ,in This is the current score matrix. This is the matrix after the previous smoothing; By constructing a multi-objective conflict scoring matrix under the functional scale, the technical effect of quantifying the intensity of conflict between different process objectives was achieved, taking into account priority and achievement. For example, on a cosmetic packaging production line with a rated capacity of 300 pieces per minute, the priority weights of three objectives—quality stability, energy efficiency, and equipment lifespan—are set as follows: , , Under the current feature projection, the comprehensive index of quality stability parameters =92, target value =95, Achievement Rate =0.968; Energy efficiency =78, target value =80 =0.975; Equipment lifespan =88, Target value =90 =0.978. Let α = 1.2, then the conflict strength between mass stability and energy efficiency is... =0.00126, similarly calculate the conflict intensity between quality stability and equipment life =0.00072, and the conflict intensity between energy efficiency and equipment life =0.00084. Construct a symmetric matrix and nullify the diagonal elements. Finally, after exponential smoothing with λ=0.3, the smoothed matrix elements are 0.00093, 0.00058, and 0.00065, respectively. This matrix serves as the functional scale input for conflict type discrimination in step S5.4, effectively reflecting the degree of weak conflict between targets. S5.4: The time-perturbation intensity index and the functional conflict scoring matrix are integrated and analyzed. A multi-scale decision tree classification algorithm is used to identify the conflict type in order to distinguish between local conflicts and global conflicts and generate conflict type labels, including time-varying perturbation type and functional target type. S5.5: Based on the influence weights between parameters in the parameter-target correlation matrix, and combined with the temporal correlation between conflict type labels and feature vectors, a path backtracking algorithm is executed to identify conflict propagation paths and generate an influence path topology map for subsequent conflict contribution ranking and actionable suggestion generation.

[0016] Step S6: Based on the conflict type label, influence path, and parameter-target correlation matrix, calculate the contribution weight ranking of each conflict source and generate a structured conflict report. The structured conflict report includes conflict type, influence path, weight ranking, and actionable recommendations. Specifically, it includes: S6.1: Based on the conflict type label, the detected local time-varying conflicts and global functional conflicts are classified and grouped. The initial intensity value of each type of conflict is normalized by the conflict intensity quantification model to obtain the conflict intensity vector. S6.2: Perform matrix dot product operation on the conflict intensity vector and the parameter-target correlation matrix to identify the influence intensity distribution of each conflict type under the corresponding process target dimension, and generate the conflict-target influence distribution matrix; Using the conflict intensity vector and the parameter-target correlation matrix as input data, matrix multiplication (parameter: element-wise multiplication mode) is employed to map the influence intensity of conflict types across different process target dimensions. Furthermore, by using a matrix dimension alignment algorithm (parameters: the number of conflict types matches the number of targets, and a zero-padding strategy is used for missing matching items), the conflict intensity vector elements are matched one-to-one with the corresponding process target association weights, ensuring a one-to-one mapping relationship between each conflict type and its affected target. Furthermore, the formula for calculating element-wise product is as follows: in, For the elements of the conflict-target influence distribution matrix, For the conflict intensity vector at the th The value for each conflict type, For the parameter-target correlation matrix in the th Line number The column's association weight value enables a weighted distribution quantification of conflict intensity and target relevance; Furthermore, a matrix normalization algorithm is applied (parameter: row-wise normalization to maintain the comparative significance between conflict types) to achieve a proportional expression of the influence intensity values ​​of each conflict type under different process objectives, so as to obtain a unified dimension when calculating the comprehensive contribution in the future; Furthermore, a sparse matrix compression storage method is adopted (parameter: the threshold is set to 5% of the maximum value of the influence intensity) to set low-influence targets to zero, reduce the subsequent computational burden and highlight the main conflict influence paths; By performing matrix dot product and normalization, the conflict intensity vector and parameter-target correlation matrix are transformed into a conflict-target influence distribution matrix that reflects the degree of influence of different conflict types on each process target dimension, thereby achieving a quantitative and visual representation of the conflict influence. For example, on a beverage packaging production line with a rated speed of 250 pieces per minute, the conflict intensity vector is calculated via step S6.1. This corresponds to local conflict types and global conflict types. The parameter-target correlation matrix is ​​as follows: The columns represent the weights associated with quality stability, energy efficiency, and equipment lifespan objectives, respectively. Element-wise multiplication is used for calculation, and the influence intensity of local conflicts under each objective is as follows: The intensity of the global conflict's impact on each objective is as follows: Perform row-by-row normalization on both rows separately. The impact of local conflicts is [percentage missing]. The global conflict impact ratio is After 5% threshold sparsification, targets not below the threshold are retained, generating a sparsified conflict-target impact distribution matrix. This matrix provides a directly usable matrix input for the comprehensive contribution calculation of S6.3, which effectively guides the priority allocation of resources between quality stability and equipment lifespan in subsequent strategy optimization. S6.3: Based on the conflict-target impact distribution matrix and combined with the preset process priority weighting factor, a weighted summation algorithm is used to calculate the comprehensive contribution score of each conflict source to form a ranking list of conflict source contributions. S6.4: Perform an impact path backtracking analysis on the conflict types in the conflict source contribution ranking list, and generate a structured conflict impact path tree map by combining conflict type labels and impact path information; S6.5: Based on the tree diagram of conflict impact paths and the ranking list of conflict source contributions, a structured conflict report is generated through a preset actionable suggestion rule engine. The structured conflict report includes conflict type labels, impact path diagrams, contribution rankings, and corresponding actionable suggestions.

[0017] Step S7: Based on the structured conflict report, determine whether the preset optimization intervention threshold is met. If so, output a parameter dynamic response threshold adjustment suggestion or a target priority reconstruction strategy for manual intervention or automatic strategy update module execution. Specifically, this includes: S7.1: Analyze the conflict type labels, impact paths, and contribution weight rankings in the structured conflict report, extract actionable suggestions for each conflict source, generate an actionable set of intervention suggestions, and provide a basis for subsequent strategy generation; S7.2: Based on the set of executable intervention suggestions, calculate the current conflict intensity index, which is obtained by weighted summation of the contribution weights of each conflict source, in order to quantify the severity of conflict in the current multi-objective optimization. S7.3: Compare the current conflict intensity index with the preset optimization intervention threshold to determine whether the strategy update mechanism is triggered, so as to determine whether it is necessary to perform parameter dynamic response threshold adjustment or target priority reconstruction operation. Based on the current conflict intensity index calculated by step S7.2, the threshold comparison algorithm is called (parameter: threshold type is static calibration value or dynamic calculation value, tolerance range is set to ±5%) to achieve accurate comparison between the conflict severity and the preset optimization intervention threshold; Furthermore, by using a double-precision floating-point comparison mechanism (parallel calculation of absolute and relative differences in comparison mode), the difference between the conflict intensity index and the intervention threshold is obtained, and a difference vector is generated for subsequent trigger logic judgment. Furthermore, a condition-triggered discrimination algorithm is adopted (parameter: the trigger condition is that any element of the difference vector is greater than 0 and the corresponding conflict type weight exceeds the global mean) to realize the start determination of the policy update mechanism and obtain the trigger flag signal; Furthermore, by using a decision mapping matrix (parameter: matrix dimension equals the number of conflict types × the number of strategy categories, element value is the trigger priority), the trigger flag signal is mapped to a specific strategy category, and the strategy flag code of the parameter dynamic response threshold adjustment category or target priority reconstruction category is obtained; By comparing and judging the current conflict intensity index, the current conflict intensity index is transformed into a strategy trigger flag and a category flag code, thereby achieving the expected technical effect of confirming the conditions and locking the execution direction of the strategy update mechanism. For example, on a beverage packaging production line rated at 250 units per minute, step S7.2 calculates the current conflict intensity index as follows: The preset optimization intervention threshold is Tolerance range ± The formula for calculating the absolute difference is: ,in As an indicator of conflict intensity, The intervention threshold is used to obtain the absolute difference value. Exceeding the tolerance limit The formula for calculating the relative difference is: The relative difference is approximately Exceeding the set limit The weight distribution data shows that the weight of local conflict type is... Higher than the global weight average The triggering conditions are met. The element value of the strategy category corresponding to this type in the mapping matrix is ​​2, pointing to the parameter dynamic response threshold adjustment category. Finally, the trigger flag signal and strategy category flag code 2 are output, and the system enters the parameter dynamic response threshold adjustment strategy execution process, effectively suppressing the risk of stacking failure caused by future batch sealing anomalies; S7.4: If the conflict intensity index is greater than the optimization intervention threshold, then the strategy classification decision is executed based on the conflict type label, the parameter dynamic response threshold adjustment suggestion is executed for local conflicts, and the target priority reconstruction strategy suggestion is executed for global conflicts, so as to generate a strategy update candidate set; S7.5: Sort the policy update candidate set by priority and encapsulate it into a standardized policy update instruction format for review and confirmation by the manual intervention module or direct loading and execution by the automatic policy update module, so as to realize the dynamic adjustment and closed-loop feedback of multi-objective optimization policies.

[0018] Step S8: After the strategy is executed, continuously monitor the operating status of the packaging production line, collect feedback data and compare it with historical data, and update the parameter configuration of the multi-scale feature encoder and functional semantic mapping module to achieve dynamic optimization and adaptive improvement of conflict modeling capabilities. Specifically, this includes: S8.1: Real-time monitoring of the packaging production line operation status after strategy execution, collection of multi-source parameter feedback data, including sealing temperature, operating speed, tension control, energy consumption indicators and ambient temperature and humidity, forming a multi-scale feedback dataset with three time granularities of minute, hour and batch to support subsequent model updates and performance evaluation; S8.2: Perform standardization and denoising on the multi-scale feedback dataset, and use the sliding window method to extract statistical features and fluctuation features at each time granularity to obtain the change pattern of parameters at different time scales after strategy execution, and generate an updated feature sample set for incremental training and comparative analysis of model parameters. S8.3: Based on the difference measure between the updated feature sample set and the historical feature sample set, calculate the drift index of the multi-scale feature space, and use the method of combining KL divergence and Mahalanobis distance to evaluate the feature distribution shift, so as to determine whether the online update mechanism of the multi-scale feature encoder is triggered. S8.4: If the feature drift exceeds the preset threshold, the parameters of the multi-scale feature encoder are fine-tuned based on the incremental learning strategy, and the encoder network weights are locally updated using an online self-supervised learning algorithm to maintain its decoupling ability to parameter change patterns under new working conditions, and to generate an updated multi-scale feature vector space. S8.5: Input the updated multi-scale feature vector into the functional semantic mapping module, perform semantic consistency correction based on the newly acquired process target labels and historical mapping relationships, and use graph attention network to optimize the weight allocation of the parameter-target association matrix to improve the accuracy and stability of target mapping and realize the incremental knowledge update of the functional semantic mapping module. S8.6: Based on the updated multi-scale feature encoder and functional semantic mapping module, the dual-scale conflict detection algorithm is rerun to generate updated conflict type labels, impact paths, and contribution weight rankings to verify the improvement effect of model update on conflict modeling capability and generate model iteration logs for system performance evaluation and subsequent strategy optimization reference.

[0019] 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.

[0020] 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 method for intelligent optimization of packaging parameters in a packaging production line, characterized in that, Includes the following steps: S1: Collect multi-source parameter data during the operation of the packaging production line, divide the time series into three time granularities: minute, hour, and batch, and form a multi-scale raw dataset. S2: Standardize and denoise the multi-scale original dataset, and use the sliding window method to extract the statistical and fluctuation features at each time granularity to obtain the basic change patterns of each parameter at different time scales. S3: Construct a multi-scale feature encoder, decouple the basic change patterns based on a self-supervised learning model, extract the independent change patterns of parameters on the time scale, and generate multi-scale feature vectors. S4: Input the multi-scale feature vector into the functional semantic mapping module, combine it with the preset process target label, map each parameter to its corresponding process target, and generate a parameter-target association matrix; S5: Based on the parameter-target correlation matrix and multi-scale feature vector, execute the dual-scale conflict detection algorithm to generate conflict type labels and influence paths; S6: Calculate the contribution weight ranking of each conflict source based on the conflict type label, the influence path and the parameter-target correlation matrix, and generate a structured conflict report; S7: Based on the structured conflict report, determine whether the preset optimization intervention threshold is met. If it is met, output a parameter dynamic response threshold adjustment suggestion or a target priority reconstruction strategy.

2. The intelligent optimization method for packaging parameters of a packaging production line according to claim 1, characterized in that, Following step S7, the following is also included: S8: After the strategy is executed, continuously monitor the operating status of the packaging production line, collect feedback data and compare it with historical data, and update the parameter configuration of the multi-scale feature encoder and functional semantic mapping module.

3. The intelligent optimization method for packaging parameters of a packaging production line according to claim 1, characterized in that, Step S1 specifically includes: Based on the industrial IoT platform, multi-source parameter data during the operation of the packaging production line are acquired to obtain the raw time-series data stream; The original time-series data stream is processed with timestamp alignment and sampling rate unification. Linear interpolation and sliding window mean filtering algorithms are used to repair missing and outlier values, generating a time-aligned structured original dataset. The structured original dataset is sliced ​​into time windows based on minute-level time granularity, and the instantaneous values ​​and statistical features of each parameter at the minute-level granularity are extracted to generate a minute-level time series data subset. Sliding window aggregation calculations are performed on the structured original dataset based on hourly time granularity. The trend features of each parameter at the hourly granularity are extracted using mean, variance and range statistical methods to generate an hourly time series data subset. The structured original dataset is divided into batches based on batch-level production cycle identifiers. The overall distribution characteristics of each parameter at the batch-level granularity are extracted by combining the start and end timestamps of the batches, and a batch-level time series data subset is generated. The minute-level, hour-level, and batch-level time series data subsets are fused and formatted to generate a multi-scale original dataset containing multiple time granularity labels.

4. The intelligent optimization method for packaging parameters of a packaging production line according to claim 3, characterized in that, In step S1, multi-source parameter data is acquired based on the industrial IoT platform. Sensors with a sampling frequency of 1Hz to 10Hz are used for real-time acquisition. The MQTT industrial message protocol (QoS level 2) is used, and NTP network clock synchronization accuracy ≤10ms is used for unified time tag processing. Through message queue buffering mechanism and CRC32 integrity verification algorithm, the multi-source parameter data is transformed into a raw time-series data stream with unified structure, comparable time, and complete verification.

5. The intelligent optimization method for packaging parameters of a packaging production line according to claim 1, characterized in that, Step S2 specifically includes: Missing values ​​are detected and imputed in the multi-dimensional parameters of the original multi-scale dataset to generate a complete time series dataset. Based on the complete time series dataset, standardization processing is performed on each parameter to generate a standardized parameter sequence. The standardized parameter sequence is denoised to obtain a denoised time series signal; Based on the denoised time series signal, the sliding window method is used to set the window length and step size for three time granularities: minute, hour, and batch, respectively, to extract local statistical features at each time granularity and generate statistical feature vectors at multiple time granularities. Further extract fluctuation features from the statistical feature vectors at each time granularity, calculate the rate of change and cumulative change of feature values ​​between adjacent windows, and generate a basic change pattern feature set.

6. The intelligent optimization method for packaging parameters of a packaging production line according to claim 5, characterized in that, Step S2 further includes using a missing value detection algorithm with a detection window length of 10 points and a threshold ratio of 1, combined with linear interpolation, boundary extrapolation, integrity verification, Z-score normalization, and wavelet threshold denoising algorithm to remove high-frequency noise. The mean, variance, kurtosis, skewness, and multi-time granularity statistical feature vectors are extracted using a sliding window method (window length is set in minutes, hours, and batches, with a step size of 30 seconds).

7. The intelligent optimization method for packaging parameters of a packaging production line according to claim 1, characterized in that, Step S3 specifically includes: Based on the aforementioned fundamental change patterns, a self-supervised learning framework is constructed. A temporal comparison learning strategy is adopted to extract unlabeled features from parameter change patterns at multiple temporal granularities, thereby constructing a potential feature space for parameters at different time scales. A scale separation operation is performed on the latent feature space, and a multi-scale wavelet packet transform is used to decompose the feature vector in the frequency domain, dividing the parameter change mode into instantaneous fluctuation components and steady-state offset components. Based on the instantaneous fluctuation component and the steady-state offset component, a scale feature encoder is designed, and a long short-term memory network is used to perform local temporal modeling of the high-frequency fluctuation component to generate an instantaneous disturbance feature vector. A trend extraction operation is performed on the steady-state offset components. The long-term evolution trend of the parameters is extracted by a combination of sliding window averaging and polynomial fitting, and a steady-state offset feature vector is generated. The instantaneous disturbance feature vector and the steady-state offset feature vector are concatenated and fused, and a fully connected neural network is used to perform nonlinear mapping to generate a unified multi-scale feature vector.

8. The intelligent optimization method for packaging parameters of a packaging production line according to claim 1, characterized in that, Step S4 specifically includes: For each dimension of the parameter feature in the multi-scale feature vector, target label alignment is performed. Based on the preset process target classification system, the parameter features are initially semantically matched with three types of process targets: quality stability, energy efficiency and equipment life, to generate a preliminary parameter-target mapping table. The initial parameter-target mapping table is semantically enhanced by using parameter-target causal rules defined by domain experts to correct and expand the mapping relationship, thereby generating an enhanced parameter-target semantic mapping table. For each parameter-target pair in the enhanced parameter-target semantic mapping table, the computational model calculates its semantic association strength and generates a parameter-target association strength matrix; Based on the parameter-target correlation strength matrix, a normalization algorithm is used to standardize the correlation strength of each parameter under different process targets, generating a standardized parameter-target correlation matrix. The standardized parameter-target association matrix is ​​subjected to sparsification. A semantic association threshold is set, and the parameter-target association weights below the threshold are reset to zero to generate a sparsified parameter-target association matrix.

9. The intelligent optimization method for packaging parameters of a packaging production line according to claim 8, characterized in that, Step S4 further includes performing target label alignment on each parameter feature of the multi-scale feature vector, selecting the best matching target using Euclidean distance and similarity normalization, completing indirect associations through process knowledge graph reasoning combined with domain expert causal rules and multi-hop reasoning, forming an enhanced parameter-target semantic mapping using a weight fusion algorithm, subsequently calculating semantic similarity for each parameter-target pair using a neural vector model, filling the parameter-target association strength matrix after min-max normalization, and sparsifying it using an association threshold.

10. The intelligent optimization method for packaging parameters of a packaging production line according to claim 1, characterized in that, Step S5 specifically includes: The parameter-target correlation matrix is ​​transformed by target dimension projection, and the multi-scale feature vectors are grouped based on the target semantic label to obtain the feature projection vectors in the target dimension. The sliding window difference analysis method is used to detect local mutations in the feature vectors at the time scale, calculate the feature difference between adjacent time windows, identify abnormal fluctuations in parameters caused by instantaneous disturbances, and generate a time disturbance intensity index. Based on the priority constraint relationship between process objectives and the objective achievement function, a multi-objective conflict scoring model under the functional scale is constructed. The conflict score is calculated on the feature projection vector under the objective dimension to generate a functional conflict scoring matrix. The time disturbance intensity index is fused and analyzed with the functional conflict scoring matrix, and a multi-scale decision tree classification algorithm is used to identify the conflict type and generate conflict type labels. Based on the influence weights between parameters in the parameter-target correlation matrix, and combined with the temporal correlation between the conflict type label and the feature vector, a path backtracking algorithm is executed to identify the conflict propagation path and generate an influence path topology map.

Images