SPC formula optimization algorithm and system based on industrial automatic production line

Abstract

The invention provides an SPC formula optimization algorithm and system based on an industrial automation production line, and the algorithm comprises the steps: obtaining a blanking line production state key process parameter, and obtaining a data set; based on the data set, constructing control parameters of a mean value control chart and a standard deviation control chart, and then implementing a Nelson anomaly judgment criterion to carry out real-time anomaly detection to obtain an abnormal data signal; a feature encoder structure is constructed, feature decomposition is carried out, and final fusion features are generated; and implementing a prevention optimization strategy. According to the method, parallel signal decomposition is carried out on all process parameters by constructing a feature encoder of time-frequency features, features are extracted from three time scales of transient features, process oscillation features and trend drift features, input data capable of comprehensively representing instantaneous, dynamic and long-term features of the production process are provided for subsequent analysis and optimization, and the method is suitable for the production process of the large-scale industrial production. Production formula optimization and team effective production duration extraction are realized.

Description

Technical Field

[0001] This invention relates to the field of industrial automated production line technology, and specifically to an SPC recipe optimization algorithm and system based on industrial automated production lines. Background Technology

[0002] With the rapid development of China's manufacturing industry, data acquisition and management of automotive parts production lines have become increasingly important. Traditional automotive parts blanking line production data acquisition systems mainly use Mitsubishi PLCs (Q-series CPUs) and Weintek touch screens to form a control system, including three main control units: uncoiler, press, and stacker.

[0003] This system employs distributed control via multiple PLCs, achieving basic communication through a fieldbus. Data acquisition focuses on a limited number of key parameters, such as SPM stroke count and uncoiler tension, which are then stored briefly in the PLCs' built-in memory. However, this approach presents several challenges: Incomplete data collection, with only a small number of key parameters (approximately 20-30 data points) collected, fails to fully reflect the production status, leading to inaccurate production efficiency analysis. In particular, the inability to analyze production line data parameters results in low SPM and low processing efficiency after the unit actually starts. Short data retention periods, coupled with the limited storage capacity of traditional PLCs, make data easily lost and unable to meet production traceability and analysis needs. This is especially true during the material handling line process, where ADC data retention is prone to errors and has short lifespans. The lack of intelligent analysis prevents automatic comparison and optimization of production parameters, relying on manual experience to adjust parameters, resulting in low production efficiency, especially since differences in process parameters among operators per shift lead to long downtimes. Insufficient real-time monitoring and a lack of effective visual monitoring methods prevent timely detection of production anomalies. Furthermore, low system integration, with isolated PLC and office networks hindering effective data sharing, results in two types of networks in the factory: a PLC network and an office network. The PLC network has two identical subnets that are not connected.

[0004] Existing technologies, such as CN120560055A, disclose a dynamic threshold optimization-based SPC control method, system, device, and medium. This method, based on a sliding window and outlier filtering dynamic threshold optimization SPC control approach, addresses the problem of poor adaptability of static thresholds. However, this method still has shortcomings in outlier detection and filtering, and further improvements in outlier detection and filtering are needed.

[0005] Specifically, while the Tukey method can identify some outliers, it is essentially a rigid filtering mechanism based on static statistical assumptions, which cannot fully adapt to the complex and ever-changing data characteristics in modern industrial automated production lines. This method relies on a fixed multiple of the interquartile range to set a uniform outlier limit, ignoring the inherent differences in data fluctuation characteristics under different production conditions and equipment operating states. During equipment startup, speed-up, or mold change phases, process parameters will exhibit reasonable transient fluctuations. If the same outlier judgment criteria as in stable production phases are used at this time, it is very easy to misjudge these normal transient data as outliers and filter them out, causing the dataset on which the control limit calculation is based to lose its integrity, and thus causing the control limit to deviate from the actual process state.

[0006] Furthermore, this method has a single dimension for identifying abnormal patterns, focusing only on the statistical distribution of data values. It cannot effectively identify time-series-based abnormal patterns (such as persistent amplitude drift, abnormal oscillations at specific frequencies, etc.). When the sensor experiences periodic interference or the transmission system has persistent abnormal friction, such anomalies cannot be effectively captured through isolated statistical tests, resulting in missed detection of anomalies.

[0007] Therefore, there is an urgent need for a management system that optimizes production formulas and improves the effective production time of refining work teams. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the purpose of this invention is to provide an SPC recipe optimization algorithm and system based on industrial automated production lines, thereby solving the problems mentioned in the background.

[0009] To achieve the above objectives, the present invention provides the following technical solution: an SPC recipe optimization algorithm based on an industrial automated production line, comprising: Key process parameters of the blanking line production status are obtained, and trend analysis methods are used to determine the trend changes in the production process and dynamically adjust the subgrouping strategy. At the same time, data is filled in based on the time decay weight interpolation method to obtain a dataset that represents the production process of the blanking line. Based on this dataset, control parameters for mean control charts and standard deviation control charts were constructed to determine the statistical boundaries and random variation ranges of normal fluctuations for each key process parameter. The Nelson anomaly detection criterion was implemented for real-time anomaly detection; when an anomaly was triggered, the anomaly data was recorded and acquired. The feature encoder structure is constructed and feature decomposition is performed. The process is as follows: For the high-frequency transient features representing the key process parameters in the abnormal data, complex Morlet wavelet transform is used to obtain the transient features. To address the process oscillation characteristics that characterize key process parameters, empirical mode decomposition (EMF) is performed: Basic parameters are set, and the signal to be decomposed is acquired and defined at a preset sampling rate. The decomposition process is executed by constructing noisy signals multiple times and calculating their residual average values. Iterative extraction of the IMF components representing the original signals of key process parameters at different frequency oscillation modes is then performed. For each IMF component, its time-domain statistical characteristics and frequency-domain energy characteristics based on Hilbert transform are calculated to form the process oscillation characteristics. To address the long-term trend characteristics of key process parameters, a deep temporal convolutional network is constructed for temporal convolutional encoding to extract trend drift features. Generate a unified feature vector ; Construct a process knowledge graph and output knowledge features that include at least the dependencies between equipment, parameters, and failure modes by building a multi-layer graph attention network. Generate final fusion features Implement preventative optimization strategies.

[0010] As a second aspect of the present invention, an SPC recipe optimization system based on an industrial automated production line is proposed, comprising: a memory and a processor, wherein the memory includes an SPC recipe optimization program based on an industrial automated production line, and when the SPC recipe optimization program based on an industrial automated production line is executed by the processor, the SPC recipe optimization of the industrial automated production line as described above is implemented.

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. To address the issue of capturing fault features hidden in the time-frequency domain structure of signals, a feature encoder based on time-frequency features is constructed to perform parallel signal decomposition on all process parameters. Features are extracted from three time scales: transient features, process oscillation features, and trend drift features. This ensures the physical clarity and discriminability of the extracted features, providing input data that can comprehensively characterize the instantaneous, dynamic, and long-term characteristics of the production process for subsequent analysis and optimization. 2. By constructing a process knowledge graph containing multiple process nodes and relational edges, this invention structurally represents the dependencies between equipment, parameters, and failure modes in the form of triples, and uses GAT to extract the contextual knowledge information most relevant to the current production state to form a knowledge feature vector, thereby enhancing the robustness of the system under complex working conditions, improving the operational stability and product quality of the blanking line, and optimizing the production formula and the effective production time of the refining team. 3. By deploying a high-performance data acquisition gateway network, complete monitoring data of 129 key process parameters, including SPM stroke count, uncoiler tension value, and feed roll length, is constructed. The dynamic adjustment strategy of SPC subgroup division and interpolation method are adopted to effectively solve the problem of data loss caused by network interruption or sensor failure. By converting the process data, which is often non-normal in actual production, into data that conforms to a normal distribution, the effectiveness of SPC control chart monitoring data is ensured. Attached Figure Description

[0012] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein: Figure 1 This is a schematic diagram of the SPC recipe optimization algorithm structure proposed in one embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the detailed data acquisition and preprocessing process proposed in one embodiment of the present invention; Figure 3 This is a schematic diagram of the SPC control chart construction and real-time monitoring process proposed in one embodiment of the present invention; Figure 4 This is a schematic diagram of the complete logical flow of implementing the prevention and optimization strategy proposed in one embodiment of the present invention. Detailed Implementation

[0013] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0014] The present invention will be further described in detail below with reference to the accompanying drawings, but this is not intended to limit the scope of the invention.

[0015] like Figure 1 - Figure 4 As shown, as a first aspect of the present invention, an SPC recipe optimization algorithm based on an industrial automated production line is proposed, comprising: Step 1: Data Acquisition and Preprocessing In this embodiment, firstly, during the data acquisition phase of the industrial automation production line, at least three high-performance acquisition gateways are deployed to build multiple industrial communication protocols. A data acquisition frequency of once per second is set to continuously acquire 129 key process parameters, thereby forming complete key process parameters for the blanking line production status. These key process parameters include at least: SPM (surges per minute) data, read from the D100 register of the Mitsubishi Q-series PLC and converted to an engineering value of 0-60 SPM using a linear transformation formula; uncoiler tension value, read from the D200 register and mapped to an actual tension value of 0-1000N using a transformation formula; and feed roller length, obtained from the D300 register, corresponding to an actual length of 0-1000mm, with measurement accuracy maintained within ±1mm.

[0016] Secondly, an improved Mann-Whitney trend analysis method is used to analyze the collected monitoring data to address the problem of traditional systems relying on manual experience to adjust parameters. Specifically, the rank sum statistic for the two time windows is calculated, and the minimum value is used to obtain the test statistic. When the p-value corresponding to the calculated U statistic is less than a preset threshold, a trend change in the production process is determined. Based on this analysis, the subgrouping strategy is dynamically adjusted: in blanking line production, the SPM (single stamping cycle) data collected for every 15 consecutive stamped parts produced constitutes a subgroup. Under normal conditions, the subgroup capacity is adjusted from the basic 15 stamping cycles to 20 cycles. This increases the subgroup capacity to improve the stability of the control chart when process fluctuations increase, avoiding excessive false alarms, while maintaining a smaller subgroup during stable periods to capture abnormal data.

[0017] Furthermore, to address the issue of missing values ​​during data acquisition, an interpolation method based on time decay weights is used for data imputation to resolve the problem of PLC data loss. In the formula, These are interpolation estimates for missing points, used to fill in the gaps in all critical process parameters. and These represent the valid observations at the time points adjacent to the missing point for any parameter currently being processed. and λ represents the time decay weight, reflecting the decay characteristics of data temporal correlation. λ is the decay coefficient, and in this embodiment, its optimal value is determined to be 0.1 through grid search optimization. Δti represents the distance between the missing time point and the known forward data point. The time difference, Δtj, is the time difference between the missing time point and the known forward data point. The time difference is expressed in seconds. k and l represent the number of data points before and after the interpolation, which are set to 5 points before and after the interpolation based on production continuity requirements.

[0018] Furthermore, to further improve the interpolation accuracy, the time series interpolated values ​​are then... As the basic input, the weighted formula is used: The final interpolated value is generated, where the spatial correlation component is... This data is obtained by spatial averaging of monitoring data for the currently missing parameters of similar equipment (such as standby stamping presses or adjacent workstations). When data for the target equipment is missing, it is supplemented by referring to the operating status of similar equipment. (Process-related component) The regression model established based on the correlation of process parameters shows that when any parameter (such as SPM strokes, tension value, or feed length) is missing, it can be estimated by other parameters that are strongly correlated with it.

[0019] Finally, to meet the requirement of SPC control charts for normal data distribution, the Anderson-Darling normality test is performed to evaluate the normality of the complete dataset for each parameter. When the test determines that the data of a certain parameter does not conform to a normal distribution, the Johnson transform function optimized specifically for that parameter is called. Using its specific transform parameters trained on historical data, the original non-normal data is transformed into a new data sequence that conforms to a normal distribution, thus obtaining accurate and reliable transformed normally distributed data.

[0020] Step 2: SPC Control Chart Construction and Real-time Monitoring In this embodiment, the core control parameters for constructing the mean control chart are based on the preprocessed data. Specifically, the centerline is first determined by calculating the average of the means of all subgroups. The subgroups are divided into 15 stamping cycles, with each subgroup representing approximately 45 seconds of production monitoring data. Then, based on the centerline calculation, the average standard deviation of all subgroups is considered. Calculate the upper control limit using the control chart coefficient (based on subgroup size n=15, determined to be 0.785 from the standard control chart coefficient table). With lower control limit This allows for the determination of the statistical boundaries of normal fluctuations in key parameters such as SPM stroke count. Understandably, this method ensures that the control chart accurately reflects the individual statistical characteristics by independently calculating the fluctuation characteristics of each key parameter.

[0021] Secondly, the standard deviation control chart (S-chart) constructed simultaneously is used to monitor process variability characteristics and ensure that the control limits can accurately reflect the inherent range of random variation in the material feeding line production process.

[0022] Furthermore, the Nelson anomaly detection criterion is implemented for real-time anomaly detection, fully encompassing the eight anomaly detection rules defined by Nelson's rules. Each rule has a corresponding sensitivity parameter to ensure timely and accurate detection of various anomaly patterns.

[0023] Example: The following is Table 1 of the mean control chart (Xbar chart) experimental data, showing some monitoring data of the SPM strokes of the blanking line within a set production cycle:

[0024] Control limit calculation results: Centerline CLM = 30.00; Upper control limit UCLM = 30.00 + 0.785 × 0.28 = 30.22; Lower control limit LCLM = 30.00 - 0.785 × 0.28 = 29.78.

[0025] The following is Table 2, which shows experimental data of the standard deviation control chart (S-chart), displaying partial monitoring data of the SPM strokes of the blanking line within a set production cycle:

[0026] Control limit calculation results: centerline CLS=0.28; upper control limit UCLS=1.58×0.28=0.44; lower control limit LCLS=0.42×0.28=0.12.

[0027] Based on the above rules, real-time analysis and judgment were performed on the SPM stroke monitoring data of the blanking line production data: Anomaly analysis indicated the following: Mean control chart anomalies: Subgroup 004: SPM stroke mean 32.61, exceeding the upper control limit of 30.22 according to rule 1, judged as an abnormally upward skewed mean; Subgroup 006: SPM stroke mean 27.31, below the lower control limit of 29.78 according to rule 1, judged as an abnormally downward skewed mean. Standard deviation control chart anomalies: Subgroup 007: Standard deviation 1.05, significantly exceeding the upper control limit of 0.44 according to rule 1, judged as an abnormal process fluctuation. These anomaly signals indicate that both the centralization and dispersion of the production process are statistically out of control, but the specific root cause of the failure is unknown.

[0028] Finally, when any abnormal condition is triggered, the abnormal data recording process is executed. The recorded content includes at least the abnormal timestamp, the abnormal parameter type (e.g., SPM strokes, uncoiler tension value, feed roller length), the triggering criterion number, the abnormal magnitude, and the complete data of the corresponding subgroup. Subsequently, the raw data of this parameter within the abnormal time period, including data for 30 minutes before and after the abnormal point and its subgroup statistical sequence data (e.g., mean, standard deviation), are packaged together into a standard CSV format data package for storage.

[0029] As an embodiment of the present invention, it should be noted that in practical applications of the prior art, it does not cover all the characteristics of production anomalies. Many key anomaly data (such as high-frequency resonance caused by bearing wear, periodic oscillation caused by transmission system loosening, low-frequency trends caused by controller parameter drift, and nonlinear effects generated by multiple parameters) are hidden in the time-frequency domain structure of each parameter signal. These characteristic signals cannot be directly obtained through distribution transformation.

[0030] Based on this, the following steps are needed to optimize feature encoding and knowledge graph analysis. For example, when early wear occurs in the bearing of an uncoiler, the feature encoding proposed in this invention is used to analyze transient features, process oscillation features, and trend drift features. Its vibration features will simultaneously manifest as periodic pulses in the transient features of multiple related parameters (such as SPM strokes, frame vibration, and motor current). It is necessary to extract the transient impact component simultaneously through feature encoding. At the same time, through the constructed knowledge graph, diagnostic basis is extracted based on the association path between "uncoiler bearing, causing, high-frequency vibration" and "high-frequency vibration, affecting, motor current" to reduce maintenance costs.

[0031] Step 3: Construct the feature encoder structure and perform feature decomposition. First, based on the received standard CSV format data packets, data integrity is verified using the CRC32 cyclic redundancy check algorithm to ensure that no damage or loss occurs during data transmission and storage. Simultaneously, a timing consistency check is performed to verify the continuity of data timestamps.

[0032] Secondly, based on the validated abnormal data, a feature encoder is constructed to decompose all data features in parallel. Production features are extracted for different time scales, with the following specific steps: S3-1. For the high-frequency transient characteristics of all key process parameters, complex Morlet wavelet transform is used to obtain the transient characteristics. : Generate analytical wavelets using wavelet basis functions; then process the original signal for each parameter. (Including 129 parameters such as SPM stroke count, uncoiler tension, and feed roll length) A continuous wavelet transform is performed to calculate wavelet coefficients, including the band adjustment parameter 'a' and the translation parameter 'b'. A smaller band adjustment parameter 'a' results in a "thinner" wavelet function, used to capture rapidly changing high-frequency components in the signal; a larger band adjustment parameter 'a' results in a "fatter" wavelet function, used to extract slowly changing low-frequency components in the signal. By continuously changing the value of 'b', the wavelet function can slide along the time axis, thus examining the local characteristics of the signal at different times. Finally, based on the wavelet coefficients at each time scale, wavelet energy is calculated for each parameter to characterize the band energy distribution, wavelet entropy, quantify the energy distribution complexity, and peak factor to detect transient impact intensity, thereby forming a unified feature vector that comprehensively characterizes high-frequency transient anomalies, including key process parameters. .

[0033] S3-2. For the oscillation characteristics of processes such as tension fluctuations in the uncoiler, feed roller vibrations, and hydraulic system pulsations, empirical mode decomposition is performed. In practice, the maximum number of iterations is set to 100, and the screening threshold is 0.3. Optimization is performed using grid search based on four consecutive months of historical production data from the blanking line to suppress mode aliasing and reduce noise impact. S3-21. Set basic parameters and define the signal to be decomposed: T=180000, 30 minutes of outlier data were collected at a sampling rate of 1kHz; initial noise amplitude The average number of trials N=100 was used to determine the optimal value based on experimental verification, and the adaptive noise figure was adopted. The typical decomposition of IMF components has K=7, and the optimal statistical analysis result is taken.

[0034] S3-22. Perform the first decomposition, K=1: For i=1 to N, construct the noisy signal: In the formula, Given Gaussian white noise with a mean of 0 and a variance of 1, for each noisy signal... Perform standard EMD decomposition to obtain the first IMF component: Calculate the average value of the first IMF component: Extract the first residual: Understandably, during the first decomposition, Gaussian white noise of a specific amplitude is added to the original oscillation signal that characterizes the key process parameters, and empirical mode decomposition is performed on each noisy signal. The first IMF component of each signal is then taken and ensembled to obtain the first IMF component representing the most dominant high-frequency oscillation mode in the original signal. Subsequently, this component is removed from the original signal, resulting in the first residual signal containing all mid- and low-frequency components and long-term trends.

[0035] Perform the Kth decomposition, K≥2: For i=1 to N, construct the noisy signal: In the formula, This indicates Gaussian white noise The (k-1)th IMF component obtained from EMD decomposition is dynamically adjusted for noise amplitude based on the previous residual fluctuation (the residual signal obtained from the previous decomposition is used as the new processing object); for each noisy signal Perform standard EMD decomposition, take the first IMF component, and calculate the average value of the kth IMF component: Calculate the k-th residual: .

[0036] When the residual When a function is monotonic or has fewer than two extreme points, the decomposition stops. Based on statistical analysis of production data from the material feeding line, a typical decomposition is obtained. The components can effectively separate oscillation modes at different time scales.

[0037] S3-23, Decomposition of each process parameter yields... Components, calculate and extract process oscillation characteristics ,include: Time-domain statistical characteristics include: By averaging all amplitude points of the IMF component over the entire signal length T, we obtain the mean characteristic that reflects the average energy level of the IMF component (SPM impulse anomaly high-frequency noise level) over the entire time period and characterizes the reference shift of the corresponding frequency component. ; By calculating the amplitude and mean at each time point The sum of squares of the differences, divided by the signal length T-1 and then square-rooted, reflects the standard deviation of the IMF component (intermediate frequency vibration intensity of unwinding tension fluctuation) at the corresponding frequency component. This is used to quantify the range of variation in its oscillation amplitude; a larger value indicates a larger oscillation amplitude for that frequency component. Skewness characteristics By analyzing the production data of the blanking line around the average The third moment and the cube of the standard deviation The ratio is obtained, reflecting the asymmetry of the distribution of the IMF component (feed roller length deviation). A positive value indicates that the distribution is skewed to the right, and a negative value indicates that it is skewed to the left. Kurtosis characteristics Through fourth moment and standard deviation The ratio of the fourth power minus the baseline value of 3 is used to measure the sharpness or flatness of the distribution pattern of production data on the material drop line compared to a normal distribution, thus effectively characterizing the prominence of the impact component in various fault signals, and the kurtosis feature. Positive values ​​indicate a sharper distribution than a normal distribution, while negative values ​​indicate a flatter distribution.

[0038] Frequency domain energy characteristics, including: The instantaneous frequency of each IMF component is calculated using the Hilbert transform. and instantaneous amplitude : In the formula, This represents the instantaneous amplitude of the i-th IMF component at time t, characterizing the instantaneous energy intensity of that IMF component (SPM impulse anomaly high-frequency impact intensity) at time t, reflecting the change in the signal envelope. This represents the instantaneous frequency of the i-th IMF component (the change in vibration frequency of the uncoiler drive system) at time t. A decrease indicates a reduction in the equipment's vibration frequency. Represents the Hilbert transform of the i-th IMF component; Marginal Spectral Energy In obtaining the instantaneous amplitude of each IMF component based on the Hilbert transform Then, by integrating its squared value over the entire signal time length T, the marginal spectral energy of the i-th IMF component (fault energy distribution in different frequency bands) is calculated. An increase in its value indicates that the fault energy in that frequency band, caused by the slow degradation of equipment performance, is continuously accumulating. Energy entropy Obtain the marginal spectral energy of each IMF component. Then, first calculate the proportion of each component energy in the total energy. Subsequently, based on the energy proportion of all components, the information entropy formula is used to calculate and quantify the degree of concentration or dispersion of abnormal production energy in different IMF components. A lower energy entropy value indicates that the fault energy is concentrated in a few specific components, reflecting the existence of prominent local fault modes; while a higher entropy value means that the energy distribution is more uniform.

[0039] Finally, based on the IMF components obtained from empirical mode decomposition, the time-domain statistical characteristics of mean, standard deviation, skewness, and kurtosis, as well as the instantaneous amplitude, instantaneous frequency, marginal spectral energy, and frequency-domain energy characteristics of energy entropy, are combined to form the process oscillation characteristic vector. .

[0040] S3-3. To address long-term trends such as feed roller length deviation, equipment performance degradation, and slow drift of process parameters, a deep temporal convolutional network is constructed for temporal convolutional encoding to extract trend drift features. : S3-31. Construct a temporal convolutional network. Let the model input be a temporal feature tensor of shape (W,d), where W=60 is the time window length, corresponding to 30 minutes of production data, and d is the feature dimension, corresponding to all 129 key process parameters.

[0041] S3-32. First, the input features are mapped to a high-dimensional space through a feature embedding layer: In the formula, The input sequence contains complete feature data for all 129 process parameters within the past 30 minutes. For embedding weight matrix; For bias terms; This is the initial hidden state.

[0042] Secondly, convolution is performed to obtain the production pattern under long-term trend drift. The calculation of features at each time scale follows the formula: In the formula, The kernel size corresponds to the short-term trend of feed roller length deviation, the medium-term trend of uncoiler tension fluctuation, and the long-term drift law of SPM stroke anomaly, respectively. The expansion coefficient increases with increasing scale index i. For activation function, It is a one-dimensional convolution.

[0043] Finally, the outputs at each time scale are weighted and fused using a channel attention mechanism, enabling the model to focus on the time scale features most important to the current production status of the material feeding line. Global average pooling is then used to compress the time scale features into a fixed-dimensional trend representation vector, generating a trend drift feature vector. This includes: the slow cumulative effect of feed roller length deviation; the gradual drift of uncoiler tension value; the long-term stability changes of SPM stroke data; and the slow trend characteristics of 126 other process parameters.

[0044] S3-4. Construct a dynamic weighting function to dynamically allocate and adjust the weights of features at each scale during the fusion process based on the real-time process status, and construct a complete scale feature vector. Each scale feature vector includes a transient feature vector representing transient impacts, a process oscillation feature vector representing periodic fluctuations, and a trend drift feature vector representing long-term drift. In the formula, These are the initial weighting coefficients, which control the baseline importance of the features. The attenuation coefficient controls the attenuation rate of historical characteristics. This is a process state function, taking values ​​(0,1), dynamically reflecting production stability. The dynamic adjustment weights for the current process stability indicators are calculated in real time based on 129 parameters. The stability threshold was determined based on four months of historical data to ensure that more attention is paid to real-time characteristics when the process is unstable, and more reliance is placed on long-term trends when the process is stable. This refers to the sensitivity coefficient. It should be noted that the production status of the blanking line refers to the overall operating condition of the production line, including whether it is operating normally, its efficiency, its stability, and whether there is any shutdown. The process status, on the other hand, refers to the real-time values ​​and fluctuations of a series of key process parameters (such as SPM strokes, uncoiler tension, feed roller length, etc.) that ensure the stability of the above production status, as well as the overall operating conditions they constitute.

[0045] Subsequently, the weights of the feature vectors at each scale are normalized using the Softmax function, and the feature vectors are then subjected to Z-score normalization. Finally, a weighted summation formula is used to generate the final unified feature vector. .

[0046] In one embodiment of the present invention, it should be noted that signal decomposition quality optimization is also required during the empirical mode decomposition process. This is because unoptimized decomposition will produce mode aliasing, causing the frequency components of different faults to become entangled, making it impossible for subsequent models to accurately trace the source. If the uncoiler bearing experiences early wear, the resulting mid-frequency vibration characteristics (200Hz) must be clearly separated from the signal; if aliasing occurs with the normal vibration (500Hz) of gear meshing, the resulting fuzzy features will not be effectively identified.

[0047] Based on the above technical concept, the specific process for optimizing signal decomposition quality is as follows: First, an envelope symmetry check is introduced to ensure that each IMF component satisfies the local symmetry requirement. For each IMF component, upper and lower envelopes are constructed based on its local extrema, and the symmetry index is calculated: In the formula, Sy represents the degree of symmetry of the envelope. and The upper and lower envelopes are constructed from the local maxima and minima using cubic spline interpolation, respectively, where T is the signal length. When the symmetry index is less than the index threshold of 0.1, the IMF component is considered to satisfy the local symmetry requirement.

[0048] Simultaneously, the system monitors mode aliasing and detects the degree of mode aliasing by calculating the correlation coefficient between different IMF components. In the formula, This represents the correlation coefficient between the i-th and j-th IMF components. and These are the mean values ​​of the corresponding IMF components, and T is the signal length. When the absolute value of the correlation coefficient between any two IMF components exceeds the absolute value threshold of 0.5, a new decomposition is performed. Gaussian white noise is added to suppress mode aliasing and ensure that the decomposition results clearly separate different frequency components.

[0049] Taking the early wear failure of the uncoiler bearing as an example: the present invention accurately captures the abnormal vibration energy rise in the 200Hz frequency band through the oscillation feature vector, the transient feature vector synchronously detects the associated micro-impact signal, and the trend drift feature vector monitors the slow growth trend of the vibration amplitude.

[0050] Step 4: Construct a process knowledge graph S4-1. Based on four months of historical production data from the material feeding line, equipment maintenance records, and expert experience, a process knowledge graph containing multiple process nodes and relational edges is constructed. In specific implementation, the knowledge graph adopts a standard triplet data structure (head entity, relation, tail entity). Entities are extracted from equipment ledgers, process parameter tables, and maintenance work orders using entity extraction algorithms, and relationships between entities are established based on process documents and expert rules using relation extraction algorithms. Finally, the graph is verified and confirmed by domain experts.

[0051] Typical examples of triples are shown in Table 3 below:

[0052] S4-2. To obtain the complex dependencies between process entities in the knowledge graph, this embodiment employs a multi-head graph attention network (GAT) to enhance the features of graph nodes, thereby enabling information propagation between nodes. It can be understood that this network, through a multi-head attention mechanism, allows each node to aggregate information from its neighboring nodes. Specifically, the calculation process of the output features of the k-th attention head for node i is as follows: The attention coefficient of the k-th attention head is calculated using the following formula: In the formula, The initial feature vectors for nodes i and j are constructed based on the physical properties (such as dimensions and normal range) and historical statistical characteristics (such as mean and variance) of process parameters. Let be the set of neighbors of node i, containing all nodes directly connected by relation edges. This is a learnable weight matrix used for linearly transforming node features. This is a learnable attention vector used to calculate the association strength between nodes. K=8 is the number of attention heads. It extracts multi-dimensional context-related knowledge from different feature subspaces. LeakyReLU is the activation function, and the negative slope is set to 0.2 to prevent gradient vanishing. It is an ELU activation function to enhance nonlinear expressive power.

[0053] It is understandable that, through a multi-layer graph attention network, after feature enhancement of all nodes is completed, the feature vector output by each node through the multi-head attention mechanism is... A global average pooling operation is performed to obtain the overall topological structure and semantic relationships of the process knowledge graph. At this point, the system outputs the knowledge features. It includes equipment, parameters, and failure modes, as well as complex dependencies with process rules.

[0054] S4-3, unify feature vectors for multi-scale data features ) and knowledge characteristics To merge: In practice, the adaptive weighting coefficients are first calculated: Then, the final fused features are generated. When the system determines that the data quality is high and the statistical features are reliable, Approaching 1, fused features tend to unify the feature vector. When the data is uncertain or noisy, but the process knowledge graph has clear rules, λ approaches 0, and the fused features tend to be more knowledge-based. : In the formula, This is a unified feature vector obtained by fusing multi-scale features from 129 process parameters. The feature vector of process knowledge extracted from the knowledge graph contains process rules and equipment association knowledge. The weight matrix for fusion gating is obtained through training on historical data. To incorporate the bias term of the gating, optimization is performed through training. The sigmoid function compresses the output to the range [0,1], generating adaptive weights. , The weighting coefficients are adaptive and dynamically adjusted based on real-time data quality.

[0055] S4-4. Define the loss function as cross-entropy loss: In the formula, For genuine anomaly labels, For based on The predicted probabilities are used to ensure the robustness of the system under complex operating conditions and improve the accuracy of anomaly diagnosis; and the fusion gating parameters are optimized simultaneously through the gradient backpropagation algorithm. and .

[0056] Understandably, continuing with the example of uncoiler bearing wear: when abnormal vibration at 200Hz is detected, a model based on data features can only determine that "an anomaly exists," but cannot pinpoint the source of the anomaly as bearing wear, gear meshing failure, or other mechanical problems. However, by querying a knowledge graph, the system can perform inference based on the following triplet relationships: The third bearing of the uncoiler generates vibration at a frequency of 200Hz. This bearing connects to the drive shaft, which in turn drives the feed roller. Bearing wear leads to increased vibration energy and is accompanied by micro-impact signals. This allows the system to predict fault propagation paths: based on the relationship between "bearing wear affecting transmission accuracy" and "decreased transmission accuracy leading to feed length deviation," the system addresses the further feed accuracy issues caused by bearing wear, thus achieving preventative maintenance.

[0057] Step 5: Implement preventive optimization strategies S5-1. Conduct future production status prediction and confidence assessment. Construct a final fusion feature This is a multi-output time-series prediction model with inputs. In practice, this model preferably uses a lightweight fully connected network or a shallow time-series network. The model outputs a probability prediction sequence for key anomaly types such as SPM stroke anomalies, uncoiler tension fluctuations, and feed roller length deviations within a future time window (e.g., the next 30 sampling points). In the formula, each It is a multi-dimensional vector representing the probability of each type of anomaly occurring. Based on this model, the confidence score C of each prediction is output synchronously through the entropy of the model's predicted probability or through a parallel confidence estimation branch. The value ranges from 0 to 1, with a higher value indicating that the model is more confident in the current prediction result.

[0058] S5-2. Based on the prediction results and confidence level assessment in S5-1, dynamically adjust the process parameters to achieve optimized control of SPC. In specific implementation, first set a prediction confidence level threshold. When the predicted probability of any anomaly type Exceeding its action threshold, and its corresponding confidence level When an optimization strategy is triggered, the action threshold is a probability threshold value set separately for each specific type of anomaly. Based on the predicted probability and type of anomalies, the hierarchical optimization strategy shown in Table 4 below is implemented to achieve precise intervention:

[0059] S5-3. Record all optimization actions and their effects (such as the actual values ​​of the adjusted parameters and subsequent SPC monitoring data), and use them as new samples to feed back into the dataset of step 1 to update the prediction model of step 5-1 and the knowledge graph of step 4, and continue to improve them.

[0060] As a second aspect of the present invention, an SPC recipe optimization system based on an industrial automated production line is proposed, comprising: a memory and a processor, wherein the memory includes an SPC recipe optimization program based on an industrial automated production line, and when the SPC recipe optimization program based on an industrial automated production line is executed by the processor, the above-mentioned SPC recipe optimization for the industrial automated production line is achieved.

[0061] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. An SPC recipe optimization algorithm based on an industrial automation production line, characterized in that: The method comprises the steps of: Obtaining key process parameters of the blanking line production state, using trend analysis method to determine the trend change of the production process and dynamically adjust the sub-group division strategy; at the same time, based on the time decay weight interpolation method, the data set representing the blanking line production process is obtained by data filling; Based on the data set, the control parameters of the mean control chart and the standard deviation control chart are constructed respectively, the statistical boundary and the random variation range of the normal fluctuation of each key process parameter are determined, the Nelson discrimination criterion is implemented for real-time anomaly detection, and the abnormal data is recorded and obtained when the anomaly is triggered, The feature encoder structure is constructed, and feature decomposition is performed, and the process is: for the high-frequency transient feature representing the key process parameter in the abnormal data, a complex Morlet wavelet transform is used to obtain the transient feature ; for the process oscillation feature representing the key process parameter, empirical mode decomposition is performed: setting the basic parameters, obtaining and defining the decomposed signal at the preset sampling rate to execute the decomposition process, constructing the noise signal multiple times and calculating the residual average value, iteratively extracting the IMF component of the original signal representing the key process parameter under different frequency oscillation mode components, calculating the time domain statistical features and frequency energy features based on Hilbert transform for each IMF component to form the process oscillation feature ; for the long-term trend feature representing the key process parameter, a deep time sequence convolution network is constructed to perform time sequence convolution coding to extract the trend drift feature ; a unified feature vector is generated ; A process knowledge graph is constructed, and a multi-layer graph attention network is constructed to output knowledge features including at least the interdependence among equipment, parameters and failure modes , generate final fusion features ; implement a prevention optimization strategy.

2. The SPC formula optimization algorithm of claim 1, wherein: Before calculating the time domain statistical features and frequency domain energy features of each IMF component, and after iteratively extracting the IMF components, the signal decomposition quality needs to be optimized, and the specific process is as follows: Symmetry detection of envelope line is performed, and the symmetry index of the upper envelope line and the lower envelope line constructed by local extreme points is calculated. When the index is less than a preset threshold, it is confirmed that the IMF component meets the local symmetry requirement. At the same time, modal aliasing monitoring is performed, and the correlation coefficient between different IMF components is calculated. When the absolute value of the correlation coefficient of any two components exceeds a preset threshold, the original signal is re-decomposed, and modal aliasing is suppressed by adding noise.

3. The SPC formula optimization algorithm of claim 1 or 2, wherein: The process of iteratively extracting the IMF components is as follows: Set the basic parameters, including the signal to be decomposed, the initial noise amplitude, the average number of times, and the adaptive noise coefficient, and obtain the typical decomposition IMF component number based on the optimal statistical analysis result; perform the iterative decomposition process: in the first decomposition, add a specific amplitude of Gaussian white noise to the original signal representing the initial oscillation data of the key process parameter, and perform empirical mode decomposition and mean calculation to obtain the first residual representing the remaining signal after removing the high-frequency component; in the subsequent Kth decomposition, the residual signal obtained by the previous decomposition is taken as the new processing object, the noise amplitude is dynamically adjusted, and the empirical mode decomposition and mean calculation are repeatedly performed, and the subsequent IMF components are extracted in turn until the remaining residual signal becomes a monotonic function or the extreme points are less than two, and the IMF components are extracted based on the statistical analysis of the blanking line production data.

4. The SPC formula optimization algorithm of claim 3, wherein: The key process parameters at least include: SPM stroke data, read from the D100 register of the Mitsubishi Q series PLC; uncoiler tension value, read from the D200 register; and feeding roller length, obtained from the D300 register; Wherein, in the forming process oscillation characteristics The time-domain statistical characteristics of each of the obtained IMF components include: a mean value characteristic representing an abnormal high-frequency noise level of the SPM impact, a standard deviation characteristic representing the intensity of the mid-frequency vibration of the unwinder tension fluctuation, a skewness characteristic representing the asymmetry of the length deviation distribution of the feeding roller, and a kurtosis characteristic representing the burst peak degree of various fault signals. The frequency domain energy features of each IMF component include: instantaneous amplitude features representing abnormal high-frequency impact strength of SPM stroke, instantaneous frequency features reflecting vibration frequency changes of the uncoiler transmission system, marginal spectrum energy features describing fault energy distribution in different frequency bands, and energy entropy features quantifying the concentration degree of production anomalies.

5. The SPC formula optimization algorithm of claim 1, wherein: Extracting trend drift features The process is: A temporal convolutional network is constructed to map the input features to a high-dimensional space through a feature embedding layer: wherein, is the input sequence; is an embedding weight matrix; is a bias term; is an initial hidden state; The production pattern under long-term trend drift is obtained by performing convolution, wherein the calculation of each time scale feature follows the formula: , wherein, is a convolution kernel size; is an expansion coefficient, is an activation function; The trend drift feature vector is generated by weighting fusion through a channel attention mechanism and using global average pooling to compress the time scale feature into a fixed dimension trend representation vector .

6. The SPC formula optimization algorithm of claim 1, wherein: generating a final unified feature vector The process is: A dynamic weight function is constructed, and the weight of each scale feature vector in the fusion process is dynamically allocated and adjusted according to the real-time process state: , wherein, is an initial weight coefficient, is a decay coefficient, is a process state function; The weight of each scale feature vector is subjected to Softmax normalization processing to ensure that the sum is 1, and each feature vector is subjected to Z-score standardization processing respectively, and a final unified feature vector is generated through a weighted summation formula .

7. The SPC recipe optimization algorithm of claim 1, wherein: Constructing a process knowledge graph and outputting knowledge features The process is: The standard triple data structure is adopted, entity extraction and relation extraction algorithms are adopted, and a process knowledge graph is constructed based on the blanking line historical production data, equipment maintenance records and expert experience; a multi-head graph attention network is used to enhance the features of the nodes of the process knowledge graph, so as to realize the information transmission between the nodes; after completing the feature enhancement of all nodes, the feature vectors output by the multi-head attention mechanism of each node are subjected to a global average pooling operation to obtain the overall topological structure and semantic association of the process knowledge graph, and output knowledge features . .

8. The SPC recipe optimization algorithm of claim 1, wherein: The preventive optimization strategy process is implemented as: A final fusion feature is constructed For the input multi-output time series prediction model, the output is a probability prediction sequence for the key abnormal types such as SPM frequency anomaly, decoiler tension fluctuation, and feeding roller length deviation within a future time window ; At the same time, based on this model, the confidence score C of each prediction is output by predicting the entropy of the model probability or through a parallel confidence estimation branch; according to the obtained prediction result and confidence evaluation score, the process parameters are dynamically adjusted to realize the optimization control of SPC; all optimization actions and their effects are recorded and used as new samples to update the prediction model and knowledge graph for continuous improvement.

9. The SPC recipe optimization algorithm of claim 1, wherein: Before the SPC control chart is constructed, a normality test is also performed to evaluate the accuracy of the data set. When the test result shows that the data does not conform to the normal distribution, a Johnson transformation process is started. A nonlinear transformation function is established based on the transformation parameters optimized from the historical production data, so as to convert the original process parameters into a new data set conforming to the normal distribution.

10. An SPC recipe optimization system based on an industrial automation production line, characterized by: comprises: A memory and a processor, wherein the memory comprises an SPC recipe optimization program based on an industrial automation production line, and the SPC recipe optimization program based on the industrial automation production line is executed by the processor to implement the SPC recipe optimization algorithm of any one of claims 1 to 9.

Images