Intelligent production scheduling early warning method and system for hardware machining

Through quantum-inspired deep self-coding network and multi-scale feature extraction technology, the problems of accuracy and response speed in hardware processing scheduling warning are solved, efficient and reliable abnormality detection and early warning are achieved, adapting to changes in the production environment, and improving production efficiency and equipment utilization.

CN118966643AActive Publication Date: 2024-11-15QINGDAO HAINAKUN TRADING CO LTD

Patent Information

Application Number
CN202410997895.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-24
Publication Date
2024-11-15
Estimated Expiration
2044-07-24

AI Technical Summary

Technical Problem

When facing a complex and changing production environment, the existing hardware processing and production scheduling early warning technology has problems such as insufficient accuracy, slow response speed, high false alarm rate, lack of adaptability and information fragmentation, and cannot effectively integrate abnormal detection. , diagnosis and early warning process.

Method used

The quantum-inspired deep autoencoding network (QDAN) is used to combine multi-scale feature extraction, quantum state attention mechanism, fractal convolution module, topological adaptive connection structure and non-equilibrium quantum annealing optimization to perform abnormal pattern recognition and diagnosis, and update online and multi-level push strategies to achieve real-time early warning.

Benefits of technology

It achieves high accuracy, rapid response, low false alarm rate and in-depth root cause analysis, can adapt to changes in the production environment, provide intelligent management and interpretability throughout the process, significantly improve production efficiency and equipment utilization.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118966643A_ABST
    Figure CN118966643A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of intelligent production scheduling early warning methods, in particular to an intelligent production scheduling early warning method and system for hardware machining, and the method comprises the following steps: obtaining real-time production scheduling operation data on a hardware machining production line; preprocessing the real-time production scheduling operation data to obtain preprocessed data; performing dynamic time window division based on the preprocessed data; performing multi-scale feature extraction on the divided time window data; performing abnormal mode recognition on the extracted multi-scale features by using a quantum inspiration depth self-encoding network; based on the abnormal mode recognition result, abnormal reason diagnosis is carried out; early warning information is generated and pushed; and the model is updated online according to feedback information, a complex abnormal mode can be accurately identified through an innovative quantum inspiration deep self-encoding network, and the accuracy, the precision rate and the recall rate of anomaly detection are greatly improved. Test results show that the accuracy rate reaches 98.7% and is far higher than that of a traditional method.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent production scheduling and early warning methods, and more specifically, to an intelligent production scheduling and early warning method for hardware processing and a system thereof. Background Art

[0002] With the rapid development of Industry 4.0 and intelligent manufacturing, the metal processing industry is facing unprecedented opportunities and challenges. Efficient and accurate production scheduling and real-time early warning systems have become key factors in improving corporate competitiveness. However, the existing production scheduling and early warning technologies still have many shortcomings when facing the complex and changeable metal processing environment.

[0003] Traditional rule-based early warning methods were once the mainstream in the industry. This method relies on expert experience to formulate fixed thresholds and rules. Although it is simple to implement, its limitations are increasingly prominent in the face of the complexity of modern hardware processing. First, fixed thresholds are difficult to adapt to the needs of different working conditions and product types, resulting in a high false alarm rate; second, the rule system is difficult to capture complex multivariate relationships and cannot identify subtle abnormal patterns; third, this method lacks adaptive capabilities and cannot be updated as the production environment changes.

[0004] In recent years, machine learning methods, such as random forests and support vector machines, have begun to be applied to the field of production scheduling and early warning. These methods have improved the accuracy and flexibility of early warning to a certain extent. However, they still have some problems: first, these methods usually require a large amount of labeled data for training, while in actual production, abnormal samples are often scarce; second, they have limited processing capabilities for high-dimensional and time-series data, making it difficult to fully utilize the rich information in the hardware processing process; finally, these methods have poor interpretability and are difficult to provide in-depth root cause analysis for abnormalities.

[0005] With the development of deep learning technology, some studies have begun to try to use methods such as LSTM autoencoders for anomaly detection. These methods perform well in processing time series data and can capture more complex patterns. However, they still face some challenges: first, deep learning models usually require a lot of computing resources, making it difficult to achieve real-time warnings; second, these models are prone to overfitting and lack robustness when facing new types of anomalies; third, the "black box" characteristics of deep learning models make the results difficult to interpret, which is not conducive to engineers' subsequent problem diagnosis and processing.

[0006] In addition, the existing methods generally have a common problem: they often separate anomaly detection, diagnosis and early warning, and lack a unified framework to integrate these key links. This leads to information fragmentation and affects the overall effectiveness of the early warning system.

[0007] In view of these deficiencies in the prior art, there is an urgent need for an intelligent production scheduling and early warning method and system for hardware processing. Summary of the invention

[0008] Aiming at the above problems, an intelligent production scheduling and early warning method and system for hardware processing are proposed.

[0009] The present invention provides an intelligent production scheduling and early warning method for hardware processing, comprising the following steps:

[0010] Acquire real-time production scheduling and operation data on the hardware processing production line; preprocess the real-time production scheduling and operation data to obtain preprocessed data; perform dynamic time window division based on the preprocessed data; perform multi-scale feature extraction on the divided time window data; use a quantum-inspired deep autoencoder network to perform abnormal pattern recognition on the extracted multi-scale features; diagnose the cause of the abnormality based on the abnormal pattern recognition results; generate and push warning information; and update the model online according to the feedback information.

[0011] Preferably, the pretreatment step comprises:

[0012] The improved Hampel filter is used to remove outliers, where the filter calculation formula is:

[0013]

[0014] Among them, x′ i is the filtered data point, x i is the original data point, X w For x i is a sliding window centered on , MAD is the median absolute deviation, and γ is an adjustable parameter.

[0015] Preferably, the dynamic time window division step comprises:

[0016] Calculate the adaptive window size W, the calculation formula is:

[0017]

[0018] Wherein ACF is the autocorrelation function, and threshold is a preset threshold; an overlapping sliding window sequence is generated based on the adaptive window size W and the preset overlap rate r.

[0019] Preferably, the multi-scale feature extraction step comprises:

[0020] Calculate the sample entropy SE(τ) of multiple time scales τ, and the calculation formula is:

[0021] Among them A m(r) is the probability of template matching in the m-dimensional phase space under tolerance r;

[0022] Apply wavelet packet transform to obtain multi-scale frequency domain features WPT i , and its calculation formula is:

[0023] WPT i =WPT(S i ,level=3,wavelet='db4')

[0024] Where S i is the input signal, level is the number of decomposition levels, and wavelet is the wavelet basis function

[0025] The morphological features MG(S) are extracted using the morphological gradient operator. i ), and its calculation formula is:

[0026]

[0027] where ⊕ and denote expansion and corrosion operations respectively, and B is the structural element.

[0028] Preferably, the quantum inspired deep autoencoding network comprises:

[0029] Quantum state attention mechanism, where the attention weight w i The calculation formula is:

[0030] w i =|α i | 2 =|<ψ|U i |ψ>| 2

[0031] Where |ψ> is the quantum state, |i> is the computational ground state, and α i is the complex amplitude, U i is a quantum gate operation; fractal convolution module, where the fractal convolution kernel F k The formula for generating is:

[0032] F k =IFS(W,k,s)

[0033] Where IFS is the iterated function system, W is the transformation matrix set, k is the number of iterations, s is the initial shape,

[0034] Topology adaptive connection structure, where the network connection strength C ij The calculation formula is:

[0035] C ij =σ(W·Φ(X i ,Xj )+b)

[0036] Where Φ is the continuous homology function, σ is the activation function, W and b are learnable parameters

[0037] Non-equilibrium quantum annealing optimization module, where the parameter updates follow the non-equilibrium quantum dynamics equations:

[0038]

[0039] Where ρ is the density matrix, H is the Hamiltonian, and L is the dissipative superoperator.

[0040] Preferably, the abnormal pattern recognition step further includes:

[0041] Calculate the anomaly score AS, the calculation formula is:

[0042] AS=log(RE)+λ·D W (X,X ref )

[0043] Where RE is the reconstruction error, D W is the Wasserstein distance, X ref is the reference normal sample, λ is the balance factor;

[0044] Based on the adaptive threshold T adaptive Determine abnormality, where T adaptive The update formula is:

[0045] T adaptive =β·T adaptive +(1-β)·(μ(AS)+γ·σ(AS))

[0046] Where μ and σ are the mean and standard deviation functions respectively, β is the smoothing factor, and γ is the sensitivity parameter.

[0047] Preferably, the abnormality cause diagnosis step includes:

[0048] Calculate feature importance φ using the SHAP method based on Shapley values i , and its calculation formula is:

[0049]

[0050] Where F is the feature set, v is the value function of the feature subset;

[0051] Based on the PC algorithm, a causal graph G(V,E) is constructed, where V is a node set and E is an edge set. A dynamic Bayesian network is used to simulate the propagation of anomalies in the causal graph, and its conditional probability distribution is:

[0052] P(Xt |Pa(X t ))=f(X t ,Pa(X t ),θ)

[0053] Where Pa(X t ) is X t The parent node set of each node X, f is the conditional probability distribution function, θ is the model parameter; i Root cause score RCS(X i ), and its calculation formula is:

[0054]

[0055] Where P(path|G,AS) is the probability of anomaly propagation through path path given the anomaly score AS.

[0056] Preferably, the step of generating and pushing warning information includes:

[0057] The comprehensive severity index SDI was calculated using the following formula:

[0058] SDI=w1·normalize(AS)+w2·max(normalize(RCS))

[0059] Where w1 and w2 are weight factors, normalize is the normalization function, AS is the anomaly score, and RCS is the root cause score;

[0060] Generates warning information in JSON format including timestamp, anomaly score, severity index, affected components, root cause and its importance and causal probability, and recommended actions;

[0061] Early warning information is pushed based on a multi-level push strategy, including:

[0062] When SDI <T low When , only logs are recorded;

[0063] When T low ≤SDI <T high When the system is running, it will be pushed to the monitoring system;

[0064] When SDI ≥ T high When the information is received, it will be immediately pushed to relevant personnel through multiple channels.

[0065] Preferably, the model online updating step comprises:

[0066] The elastic weight merging algorithm based on the Fisher information matrix is ​​used for incremental learning, and its loss function L(θ) is:

[0067]

[0068] Where L new is the loss function for new data, F i is the diagonal element of the Fisher information matrix, θ * The ADWIN algorithm is used to detect concept drift for the old model parameters, where for any partition (W0, W1) of the sliding window W, if it satisfies:

[0069] |μ(W0)-μ(W1)|>∈ cut

[0070] Then concept drift is detected, where m is the window size, δ is the confidence parameter;

[0071] Based on the concept drift detection results, the neural architecture search technology is used to dynamically adjust the model structure, and its optimization goals are:

[0072]

[0073] Where a is the architecture, w is the weight, R is the reward function, and π is the architecture distribution strategy. The improved knowledge distillation technology is used to transfer the knowledge of the old model to the new model, and its loss function L KD for:

[0074] L KD =α·L CE (y,σ(z s / T))+(1-α)·L KL (α(z t / T),σ(z s / T))

[0075] Where L CE is the cross entropy loss, L KL is the KL divergence, z t and z s are the logits of the teacher and student models respectively, T is the temperature parameter, and α is the balance factor.

[0076] An intelligent production scheduling and early warning system for hardware processing that implements the method comprises:

[0077] A data acquisition module is used to obtain real-time production scheduling and operation data on the hardware processing production line;

[0078] A preprocessing module, used for preprocessing the real-time production scheduling operation data;

[0079] A time window division module, used for performing the dynamic time window division;

[0080] A feature extraction module, used for performing the multi-scale feature extraction;

[0081] An anomaly recognition module, used for performing anomaly pattern recognition using the quantum-inspired deep autoencoder network;

[0082] A diagnosis module, used for diagnosing the cause of the abnormality;

[0083] An early warning module, used to generate and push the early warning information; and

[0084] An updating module is used to update the model online.

[0085] The main beneficial effects of the present invention include:

[0086] 1. High accuracy: Through the innovative quantum-inspired deep autoencoder network (QDAN), the present invention can accurately identify complex abnormal patterns, greatly improving the accuracy, precision and recall of abnormality detection. Test results show that the accuracy of the present invention reaches 98.7%, which is much higher than traditional methods.

[0087] 2. Fast response: Thanks to the parallel computing characteristics of QDAN and the non-equilibrium quantum annealing optimization algorithm, the present invention can complete anomaly detection in a very short time. The average detection time is only 12.3 seconds, which is significantly better than other methods and provides the possibility of timely prevention of production problems.

[0088] 3. Low false alarm rate: The false alarm rate of the present invention is only 1.2%, which greatly reduces unnecessary production interruptions and improves system reliability and production efficiency.

[0089] 4. In-depth root cause analysis: The diagnostic module integrates the SHAP method, cause-effect diagram construction and dynamic Bayesian network, enabling the present invention to provide accurate root cause analysis, with an accuracy rate of 93.8%, which provides strong support for engineers to quickly solve problems.

[0090] 5. Adaptability: Through online update modules, including incremental learning, concept drift detection, and model structure adaptation, the present invention can continuously learn and optimize, adapt to dynamically changing production environments, and maintain long-term high performance.

[0091] 6. Multi-scale analysis: The use of multi-scale feature extraction technology enables the present invention to capture abnormal patterns at different time and frequency scales, thereby improving the comprehensiveness and robustness of the system.

[0092] 7. Overall framework: The present invention provides a unified framework that organically combines anomaly detection, diagnosis and early warning, and realizes intelligent management of the entire process from data collection to early warning push.

[0093] 8. Interpretability: Through the quantum state attention mechanism and topological adaptive connection structure, the present invention provides the interpretability of model decisions while maintaining high performance, which helps users understand and trust the output of the system.

[0094] In summary, the present invention not only solves the problems of insufficient accuracy, slow response speed, high false alarm rate, lack of adaptive ability, etc. in the prior art, but also provides a comprehensive, efficient and reliable intelligent production scheduling and early warning solution for hardware processing. This will significantly improve the production efficiency, product quality and equipment utilization of hardware processing enterprises, and provide strong support for the intelligent transformation of the industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] Figure 1 The present invention is a flow chart of an intelligent production scheduling and early warning method for hardware processing.

[0096] Figure 2 This is a schematic diagram of the framework of an intelligent production scheduling and early warning system for hardware processing according to the present invention. DETAILED DESCRIPTION

[0097] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation method, structure, features and effects of the intelligent production scheduling and early warning method for hardware processing proposed by the present invention are described in detail below in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0098] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0099] The technical solution of the present invention is clearly and completely described below in conjunction with specific implementation methods.

[0100] In modern hardware processing factories, various equipment on the production line continuously generates a large amount of real-time data. These data contain key information such as equipment operation status, production progress, product quality, etc. The intelligent production scheduling early warning method and system proposed in the present invention are intended to make full use of these data, timely discover potential production anomalies, and provide accurate diagnosis and early warning, thereby improving production efficiency and reducing failure risks.

[0101] like Figure 1-2As shown, the hardware processing intelligent production scheduling warning system of the present invention includes a data acquisition module 1, a preprocessing module 2, a time window division module 3, a feature extraction module 4, an abnormality recognition module 5, a diagnosis module 6, a warning module 7 and an update module 8. These modules work together to realize the complete process from data acquisition to warning push.

[0102] Specifically, the intelligent production scheduling and early warning method for hardware processing of the present invention comprises the following steps:

[0103] 1. Data collection, data collection module 1 is responsible for obtaining real-time production scheduling and operation data on the hardware processing production line. In a typical hardware processing workshop, we may collect the following types of data:

[0104] Equipment parameters: such as spindle speed, feed speed, tool temperature, etc. of CNC machine tools.

[0105] Production progress: such as the processing time of each workpiece, the completion rate of the current batch, etc.

[0106] Quality indicators: such as workpiece size deviation, surface roughness, etc.

[0107] This data is usually collected by sensors distributed throughout the production line, such as ABB Field Sensors series sensors. The frequency of data collection may vary depending on the parameter type, usually between 0.1 seconds and 1 minute.

[0108] 2. Data preprocessing,The preprocessing module 2 cleans and standardizes the collected raw data. This step is crucial because the raw data may contain noise, outliers or missing values, which may affect the accuracy of subsequent analysis.

[0109] The improved Hampel filter is used to remove outliers, where the filter calculation formula is:

[0110]

[0111] Among them, x′ i is the filtered data point, x i is the original data point, X w For x i The sliding window is centered on , MAD is the median absolute deviation, and γ is an adjustable parameter, which is usually 3.

[0112] For example, suppose we are processing spindle speed data for a CNC machine. Under normal circumstances, the speed may fluctuate between 5000-6000RPM. If a 9000RPM reading suddenly appears, the Hampel filter will identify it as an outlier and replace it with the median of the surrounding data.

[0113] 3. Dynamic time window division,The time window division module 3 is responsible for dividing the continuous data stream into a series of overlapping time windows.,The purpose of this step is to capture the temporal correlation of the data and provide a basis for subsequent feature extraction.

[0114] The present invention adopts a method of adaptive window size, and its calculation formula is:

[0115]

[0116] ACF is an autocorrelation function, threshold is a preset threshold, which is usually 0.2, and an overlapping sliding window sequence is generated based on the adaptive window size W and the preset overlap rate r.

[0117] For example, for spindle speed data, if it is found that the data has a periodic change every 100 seconds or so, the window size may be set to about 100 seconds.

[0118] After that, the system will use the preset overlap rate r (usually 0.5) to generate a sequence of overlapping sliding windows:

[0119] S i =[F t ,F t+1 ,...,F t+W-1 ],t=i×W×(1-r)

[0120] This way, adjacent windows will have 50% data overlap, which helps capture patterns across windows.

[0121] 4. Multi-scale feature extraction, the feature extraction module 4 performs multi-scale feature extraction on the data of each time window. This step aims to extract features that can effectively characterize the system state from the original data.

[0122] The present invention adopts three feature extraction methods

[0123] 4.1 Sample Entropy, calculate the sample entropy SE(τ) of multiple time scales τ, and its calculation formula is:

[0124]

[0125] Among them, A m (r) is the probability of template matching in the m-dimensional phase space under tolerance r; τ is usually [1, 5, 10]; m is usually 2; r is usually 0.2 times the standard deviation;

[0126] Sample entropy can reflect the complexity of the time series. For example, for spindle speed data, if the sample entropy suddenly increases, it may mean that the equipment operation becomes unstable.

[0127] 4.2 Wavelet Packet Transform Wavelet packet transform is used to obtain multi-scale frequency domain features:

[0128] WPT i =WPT(S i ,level=3,wavelet='db4')

[0129] Among them, S i is the input signal, level is the number of decomposition levels, here it is 3, wavelet is the wavelet basis function, here we use Daubechies4 wavelet, wavelet packet transform can provide the energy distribution information of the signal in different frequency bands. For example, if the energy of a certain frequency band suddenly increases, it may mean that the equipment has abnormal vibration.

[0130] 4.3 Morphological gradient, use the morphological gradient operator to extract morphological features:

[0131]

[0132] where ⊕ and denote expansion and corrosion operations respectively, and B is the structural element.

[0133] The morphological gradient can detect rapid changes in the signal. For example, if there is a sudden jump in the dimensional deviation data of the workpiece, the morphological gradient will have a significant response.

[0134] 5. Abnormal pattern recognition, the abnormal recognition module 5 is the core of the present invention, which uses an innovative quantum-inspired deep autoencoder network (QDAN) to recognize abnormal patterns. This network combines the concepts of quantum computing, fractal geometry and topology, and can effectively capture complex abnormal patterns.

[0135] 5.1 Quantum State Attention Mechanism,First, QDAN uses the quantum state attention mechanism to assign weights to different features.,Attention weight w i The calculation formula is:

[0136] w i =|α i | 2 =|<ψ|U i |ψ>| 2

[0137] Among them, |ψ> is the quantum state; |i> is the computational ground state; α i is the complex amplitude; U i It is a quantum gate operation;

[0138] In practical applications, we can map different features (such as spindle speed, feed speed, etc.) to different quantum ground states. iThe attention weight is dynamically adjusted based on the importance of the feature. For example, if the spindle speed is found to be particularly important for anomaly detection, the system will automatically increase its corresponding attention weight.

[0139] 5.2 Fractal Convolution Module,Next, QDAN uses the fractal convolution module to process multi-scale features.,Fractal convolution kernel F k The formula for generating is:

[0140] F k =IFS(W,k,s)

[0141] IFS is an iterated function system; W is a set of transformation matrices; k is the number of iterations, usually 3-5; s is the initial shape, which can be a simple geometric shape such as a triangle;

[0142] Fractal convolution can capture self-similarity in data. For example, during processing, some abnormal patterns may present similar structures at different time scales, and fractal convolution is particularly suitable for capturing such patterns. Output feature map G k The calculation formula is:

[0143] G k =σ(F k *X+b k )

[0144] Among them, σ is the activation function, GELU is used here, * represents the convolution operation, X is the input feature, b k is the bias term.

[0145] 5.3 Topological Adaptive Connection Structure QDAN also introduces an adaptive connection structure based on algebraic topology. First, define the continuous homology function Φ:

[0146] Φ(X)={H k (X ε )|ε∈R + ,k∈Z +}

[0147] Among them, H k is the homology group of order k; X ε is the ε-neighborhood complex of X;

[0148] Network connection strength C i j is dynamically adjusted according to the continuous coherence feature:

[0149] C i j=σ(W·Φ(X i ,X j )+b)

[0150] Among them, W and b are learnable parameters, and this structure can adaptively capture the topological characteristics of the data. For example, if certain abnormal patterns form specific topological structures (such as rings or holes) in the feature space, this connection method can better identify these patterns.

[0151] 5.4 Non-equilibrium quantum annealing optimization QDAN uses non-equilibrium quantum annealing algorithm to optimize network parameters. First, define Hamiltonian H:

[0152]

[0153] Among them, σ z and σ x is the Pauli matrix; J ij is the spin coupling strength; Г is the transverse field strength.

[0154] The parameter update follows non-equilibrium quantum dynamics:

[0155] Among them, ρ is the density matrix and L is the dissipative super operator. This optimization method can effectively avoid falling into local optimality and improve the generalization ability of the model.

[0156] 5.5 Anomaly Score Calculation QDAN calculates the anomaly score AS:

[0157] AS=log(RE)+λ·D W (X,X ref )

[0158] Where RE is the reconstruction error, D W is the Wasserstein distance, X ref is the reference normal sample, λ is the balance factor, usually 0.5-1.0,

[0159] 5.6 Adaptive Threshold System Using Adaptive Threshold T adaptive To determine the exception:

[0160] T adaptive =βT adaptive +(1-β)·(μ(AS)+γ·σ(AS))

[0161] Among them, μ and σ are the mean and standard deviation functions respectively, β is the smoothing factor, usually 0.9-0.99, and γ is the sensitivity parameter, usually 2-3. If AS>T a Adaptive is judged as abnormal.

[0162] 6. Abnormality cause diagnosis,The diagnosis module 6 is responsible for analyzing the detected abnormality and finding out the possible cause.

[0163] 6.1 Feature Importance Calculation, using the SHAP method based on Shapley value to calculate feature importance φ i :

[0164]

[0165] Where F is the feature set and v is the value function of the feature subset; for example, if the SHAP value shows that the spindle speed contributes the most to a certain anomaly, then there may be a problem with the spindle.

[0166] 6.2 Causal Graph Construction The PC algorithm is used to construct a causal graph G(V,E), where V is a node set and E is an edge set. This helps to understand the causal relationship between different parameters.

[0167] 6.3 Anomaly propagation analysis, using dynamic Bayesian network to simulate the propagation of anomalies in the causal graph:

[0168] P(X t |Pa(X t ))=f(X t ,Pa(X t ),θ)

[0169] Among them, Pa(X t ) is X t The parent node set of ; f is the conditional probability distribution function; θ is the model parameter.

[0170] 6.4 Root cause score calculation, calculate each node X i Root cause score RCS(X i ):

[0171]

[0172] Where P(path|G,AS) is the probability that an anomaly propagates through the path path given the anomaly score AS.

[0173] Based on the PC algorithm, a causal graph G(V,E) is constructed, where V is a node set and E is an edge set. A dynamic Bayesian network is used to simulate the propagation of anomalies in the causal graph, and its conditional probability distribution is:

[0174] P(X t |Pa(X t ))=f(X t ,Pa(X t ),θ)

[0175] Where Pa(X t ) is X t The parent node set of each node X, f is the conditional probability distribution function, θ is the model parameter; i Root cause score RCS(Xi ), and its calculation formula is:

[0176]

[0177] Where P(path|G,AS) is the probability of anomaly propagation through path path given the anomaly score AS.

[0178] 7. Generation and push of warning information,The warning module 7 is responsible for generating warning information and pushing it according to the severity.

[0179] 7.1 Calculation of comprehensive severity index

[0180] Calculate the comprehensive severity index SDI:

[0181] SDI=w1·normalize(AS)+w2·max(normalize(RCS))

[0182] Where w1 and w2 are weight factors, usually w1 = 0.6, w2 = 0.4; normalize is a normalization function that maps values ​​to the [0,1] interval; AS is the anomaly score and RCS is the root cause score.

[0183] 7.2 Warning information is structured. The system generates warning information in JSON format, which contains the following fields

[0184] timestamp: current timestamp; anomaly_score: anomaly score AS; severity_index: comprehensive severity index SDI; affected_components: affected component list; root_causes: contains the feature name and importance of each possible root cause and causal probability; recommended_actions: a list of recommended processing actions.

[0185] 7.3 Multi-level push strategy, based on SDI value, the system adopts a multi-level push strategy

[0186] When SDI <T low (such as 0.3), only logs are recorded

[0187] When T low ≤SDI <T high (such as 0.7), push to the monitoring system

[0188] When SDI ≥ T high When the information is received, it will be immediately pushed to relevant personnel through multiple channels (SMS, email, app notification).

[0189] 8. Model online update,The update module 8 is responsible for continuously optimizing the model to adapt to the dynamic changes in the production environment.

[0190] 8.1 Incremental learning, using the elastic weight consolidation (EWC) algorithm based on the Fisher information matrix for incremental learning. The loss function L(θ) is:

[0191]

[0192] Where L new is the loss function for new data; F i is the diagonal element of the Fisher information matrix; θ * is the old model parameter; λ is the regularization coefficient, usually 0.1-18.2

[0193] 8.2 Concept drift detection, using ADWIN (ADaptiveWINdowing) algorithm to detect changes in data distribution. For any partition of sliding window W (W0, W1), if:

[0194] |μ(W0)-μ(W1)|>ε cut

[0195] Then concept drift is detected. Among them:

[0196] m is the window size; δ is the confidence parameter, usually 0.05; n is the total number of samples.

[0197] 8.3 Model structure adaptation: Based on the concept drift detection results, the neural architecture search (NAS) technology is used to dynamically adjust the model structure. The optimization objectives are:

[0198]

[0199] Where a is the architecture, w is the weight, R is the reward function, and π is the architecture distribution strategy

[0200] 8.4 Knowledge distillation: Apply improved knowledge distillation technology to transfer the knowledge of the old model to the new model. The loss function L KD for:

[0201]

[0202] Among them, L CE is the cross entropy loss, L KL is the KL divergence, z t and z S are the logits of the teacher and student models respectively, T is the temperature parameter, usually 2-5, and α is the balance factor, usually 0.5-0.7.

[0203] The hardware processing intelligent production scheduling and early warning system of the present invention is a complex system integrating multiple functional modules, aiming to achieve real-time monitoring, abnormality detection and early warning of hardware processing production lines. The core components of the system include:

[0204] Data acquisition module 1: This module is responsible for real-time data collection from various sensors and equipment on the hardware processing production line. It uses a high-precision sensor network (such as the ABB FieldSensors series) to collect data such as equipment parameters, production progress and quality indicators. The data collection frequency is dynamically adjusted according to different parameter types to ensure that key information is captured without generating too much redundant data.

[0205] Preprocessing Module 2: Preprocessing Module 2 receives the raw data and cleans and standardizes it. It uses a modified Hampel filter to remove outliers to ensure data quality for subsequent analysis. This module is also responsible for handling missing values, which may use methods such as interpolation or forward filling.

[0206] Time Window Partitioning Module 3: This module divides the continuous data stream into a series of overlapping time windows. It uses an adaptive window size method to dynamically adjust the window size according to the autocorrelation of the data to capture the temporal correlation of the data.

[0207] Feature extraction module 4: Feature extraction module 4 performs multi-scale feature extraction on the data of each time window. It uses a combination of methods such as sample entropy, wavelet packet transform and morphological gradient to extract features that can characterize the system state from different angles.

[0208] Anomaly Identification Module 5: This is the core module of the system, which uses the innovative quantum-inspired deep autoencoder network (QDAN) for anomaly pattern recognition. QDAN combines the concepts of quantum computing, fractal geometry, and topology to effectively capture complex anomaly patterns. This module also contains an adaptive threshold mechanism that can dynamically adjust the anomaly judgment criteria based on the current production status.

[0209] Diagnosis Module 6: Diagnosis Module 6 is responsible for analyzing the detected anomalies and finding possible causes. It uses the SHAP method to calculate feature importance, builds a causal graph, and uses a dynamic Bayesian network to simulate anomaly propagation, and finally calculates the root cause score of each possible cause.

[0210] Early warning module 7: Early warning module 7 is responsible for generating early warning information and pushing it according to the severity. It calculates the comprehensive severity index, generates structured early warning information, and sends the information to relevant personnel or systems according to the multi-level push strategy.

[0211] Update Module 8: Update Module 8 is responsible for continuously optimizing the model to adapt to the dynamic changes in the production environment. It uses techniques such as incremental learning, concept drift detection, model structure adaptation, and knowledge distillation to ensure that the system can remain efficient and accurate over time. These modules work closely together to form a closed-loop intelligent early warning system.

[0212] In order to verify the superiority of the present invention, we designed an embodiment and three comparative examples, and carried out detailed comparative experiments.

[0213] Embodiment 1: Intelligent production scheduling and early warning method for hardware processing of the present invention

[0214] Comparative Example 1: Traditional rule-based early warning method

[0215] Comparative Example 2: Simple Machine Learning Method (Random Forest)

[0216] Comparative Example 3: Ordinary Deep Learning Method (LSTM Autoencoder)

[0217] Test environment: We conducted a three-month test in a medium-sized hardware processing factory. The factory has 20 CNC machine tools that run 24 hours a day. We collected multi-dimensional data including spindle speed, feed speed, tool temperature, processing time, product dimensional deviation, etc.

[0218] Test indicators:

[0219] 1. Anomaly detection accuracy: the ratio of the number of correctly identified samples to the total number of samples.

[0220] 2. Precision: The ratio of the number of correctly identified abnormal samples to the number of all samples identified as abnormal.

[0221] 3. Recall: The ratio of the number of correctly identified abnormal samples to the number of all actual abnormal samples.

[0222] 4. F1 score: the harmonic mean of precision and recall.

[0223] 5. Average Detection Time (ADT): The average time from when an anomaly occurs to when it is detected by the system.

[0224] 6. False Alarm Rate (FAR): The proportion of normal samples that are mistakenly identified as abnormal.

[0225] 7. Root Cause Accuracy (RCA): The proportion of correctly identified root causes of anomalies.

[0226] We used the cross-validation method and divided the dataset into a training set (70%) and a test set (30%). For each method, we performed 10 cross-validations and took the average as the final result.

[0227] The test results are shown in the following table:

[0228] index Example 1 Comparative Example 1 Comparative Example 2 Comparative Example 3 Accuracy (%) 98.7 85.3 92.1 95.4 Accuracy (%) 97.5 78.6 89.3 93.2 Recall rate (%) 96.9 72.4 86.7 91.8 F1 score 0.972 0.754 0.880 0.925 Average detection time(s) 12.3 87.5 45.2 28.7 False alarm rate (%) 1.2 8.7 4.5 2.8 Root cause analysis accuracy (%) 93.8 61.2 78.5 85.1

[0229] According to the test results, the intelligent production scheduling and early warning method for hardware processing (Example 1) of the present invention is significantly better than other comparison methods in all indicators, and can therefore be considered as the best embodiment.

[0230] The results are analyzed and interpreted as follows:

[0231] 1. Anomaly detection performance: The present invention has achieved the best results in accuracy, precision, recall and F1 score. This is mainly due to the innovative design of the quantum-inspired deep autoencoder network (QDAN). The quantum state attention mechanism can effectively capture the complex relationship between features, the fractal convolution module can identify multi-scale abnormal patterns, and the topological adaptive connection structure enhances the model's ability to adapt to changes in data structure.

[0232] 2. Detection speed: The average detection time of the present invention is only 12.3 seconds, which is much faster than other methods. This fast response capability is derived from the parallel computing characteristics of QDAN and the non-equilibrium quantum annealing optimization algorithm, which enables the model to complete the identification of abnormal patterns in a very short time.

[0233] 3. False alarm rate: The false alarm rate of the present invention is only 1.2%, which is much lower than other methods. This shows that QDAN can not only accurately identify anomalies, but also effectively distinguish between normal fluctuations and true anomalies, which is crucial to reducing unnecessary production interruptions.

[0234] 4. Root cause analysis accuracy: The present invention also significantly outperforms other methods in root cause analysis accuracy, reaching 93.8%. This is mainly due to the system's integrated SHAP method, causal graph construction, and dynamic Bayesian network, which together form a powerful anomaly diagnosis framework.

[0235] 5. Adaptability and robustness: Although not directly reflected in the table, during the three-month test, we observed that the proposed method can adapt well to the dynamic changes of the production environment. This is attributed to the online update module of the system, including incremental learning, concept drift detection, and model structure adaptation.

[0236] The hardware processing intelligent production scheduling and early warning method of the present invention is significantly superior to existing methods in terms of accuracy, speed, reliability and interpretability of anomaly detection. It can not only quickly and accurately identify anomalies, but also provide in-depth root cause analysis while maintaining a low false alarm rate. These advantages enable the present method to provide a more reliable and efficient intelligent early warning solution for the hardware processing industry, help improve production efficiency, reduce equipment failures and quality problems, and thus create significant economic value for enterprises.

[0237] In addition, the adaptive ability of the method of the present invention also ensures its long-term application in complex and changing industrial environments. As production conditions change, the system can continuously learn and optimize to maintain a high level of performance. This ability to continuously improve is not available in traditional methods and is a key advantage in the context of Industry 4.0 and smart manufacturing.

[0238] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An intelligent production scheduling and early warning method for hardware processing, It is characterized in that The following steps are involved: Acquire real-time production scheduling and operation data on the hardware processing production line; pre-process the real-time production scheduling and operation data to obtain pre-processed data; Based on the preprocessed data, dynamic time window division is performed; and multi-scale feature extraction is performed on the divided time window data; Abnormal pattern recognition of extracted multi-scale features using quantum-inspired deep autoencoder networks; Based on the abnormal pattern recognition result, diagnose the cause of the abnormality; Generate and push warning information; And update the model online based on the feedback information.

2. The method according to claim 1, It is characterized in that The pre-processing step comprises: The improved Hampel filter is used to remove outliers, where the filter calculation formula is: Among them, x′ i is the filtered data point, x i is the original data point, X w For x i is a sliding window centered on , MAD is the median absolute deviation, and γ is an adjustable parameter.

3. The method according to claim 1, It is characterized in that The dynamic time window division step comprises: Calculate the adaptive window size W, the calculation formula is: Wherein ACF is the autocorrelation function, and threshold is a preset threshold; an overlapping sliding window sequence is generated based on the adaptive window size W and the preset overlap rate r.

4. The method according to claim 1, It is characterized in that The multi-scale feature extraction step comprises: Calculate the sample entropy SE(τ) of multiple time scales τ, and the calculation formula is: Among them A m (r) is the probability of template matching in the m-dimensional phase space under tolerance r; Apply wavelet packet transform to obtain multi-scale frequency domain features WPT i , and its calculation formula is: WPT i =WPT(S i ,level=3,wavelet=′db4′) Where S i is the input signal, level is the number of decomposition levels, and wavelet is the wavelet basis function The morphological features MG(S) are extracted using the morphological gradient operator. i ), and its calculation formula is: where ⊕ and denote expansion and corrosion operations respectively, and B is the structural element.

5. The method according to claim 1, It is characterized in that The quantum-inspired deep autoencoder network comprises: Quantum state attention mechanism, where the attention weight w i The calculation formula is: w i =|a i | 2 =|<ψ|U i |ψ>| 2 Where |ψ> is the quantum state, |i> is the computational ground state, and α i is the complex amplitude, U i is a quantum gate operation; fractal convolution module, where the fractal convolution kernel F k The generation formula is: F k =IFS(W,k,s) Where IFS is the iterated function system, W is the transformation matrix set, k is the number of iterations, s is the initial shape, Topology adaptive connection structure, where the network connection strength C ij The calculation formula is: C ij =σ(W·Φ(X i ,X j )+b) Where Φ is the continuous homology function, σ is the activation function, W and b are learnable parameters Non-equilibrium quantum annealing optimization module, where the parameter updates follow the non-equilibrium quantum dynamics equations: Where ρ is the density matrix, H is the Hamiltonian, and L is the dissipative superoperator.

6. The method according to claim 5, It is characterized in that The abnormal pattern recognition step also includes: Calculate the anomaly score AS, the calculation formula is: AS=log(RE)+λ·D W (X,X ref ) Where RE is the reconstruction error, D W is the Wasserstein distance, X ref is the reference normal sample, λ is the balance factor; Based on the adaptive threshold T adaptive Determine abnormality, where T adaotuve The update formula is: T adaptive =β·T adaptive +(1-β)·(μ(AS)+γ·σ(AS)) Where μ and σ are the mean and standard deviation functions respectively, β is the smoothing factor, and γ is the sensitivity parameter.

7. The method according to claim 1, It is characterized in that The abnormality cause diagnosis step comprises: Calculate feature importance φ using the SHAP method based on Shapley values i , and its calculation formula is: Where F is the feature set, v is the value function of the feature subset; Based on the PC algorithm, a causal graph G(V,E) is constructed, where V is a node set and E is an edge set. A dynamic Bayesian network is used to simulate the propagation of anomalies in the causal graph, and its conditional probability distribution is: P(X t |Pa(X t ))=f(X t ,Pa(X t ),θ) Where Pa(X t ) is X t The parent node set of each node X, f is the conditional probability distribution function, θ is the model parameter; i Root cause score RCS(X i ), and its calculation formula is: Where P(path|G,AS) is the probability of anomaly propagation through path path given the anomaly score AS.

8. The method according to claim 1, It is characterized in that The steps of generating and pushing warning information include: The comprehensive severity index SDI was calculated using the following formula: SDI=w 1 ·normalize(AS)+w 2 ·max(normalize(RCS)) where w 1 and w 2 is the weight factor, normalize is the normalization function, AS is the anomaly score, and RCS is the root cause score; Generates warning information in JSON format including timestamp, anomaly score, severity index, affected components, root cause and its importance and causal probability, and recommended actions; Early warning information is pushed based on a multi-level push strategy, including: When SDI <T low When , only logs are recorded; When T low ≤SDI <T high When the system is running, it will be pushed to the monitoring system; When SDI ≥ T high When the information is received, it will be immediately pushed to relevant personnel through multiple channels.

9. The method according to claim 1, It is characterized in that The model online updating step comprises: The elastic weight merging algorithm based on the Fisher information matrix is ​​used for incremental learning, and its loss function L(θ) is: Where L new is the loss function for new data, F i is the diagonal element of the Fisher information matrix, θ * The ADWIN algorithm is used to detect concept drift for the old model parameters, where for any partition of the sliding window W (W 0 ,W 1 ), if: |μ(W 0 )-μ(W 1 )|>∈ cut Then concept drift is detected, where m is the window size, δ is the confidence parameter; Based on the concept drift detection results, the neural architecture search technology is used to dynamically adjust the model structure, and its optimization goals are: Where a is the architecture, w is the weight, R is the reward function, and π is the architecture distribution strategy. The improved knowledge distillation technology is used to transfer the knowledge of the old model to the new model, and its loss function L KD for: L KD =α·L CE (y,σ(z s / T))+(1-a)·L KL (σ(z t / T),σ(z s / T)) Where L CE is the cross entropy loss, L KL is the KL divergence, z t and z s are the logits of the teacher and student models respectively, T is the temperature parameter, and α is the balance factor.

10. An intelligent production scheduling and early warning system for hardware processing that implements the method described in any one of claims 1 to 9, It is characterized in that include: A data acquisition module is used to obtain real-time production scheduling and operation data on the hardware processing production line; A preprocessing module, used for preprocessing the real-time production scheduling operation data; A time window division module, used for performing the dynamic time window division; A feature extraction module, used for performing the multi-scale feature extraction; An anomaly recognition module, used for performing anomaly pattern recognition using the quantum-inspired deep autoencoder network; A diagnosis module, used for diagnosing the cause of the abnormality; An early warning module, used to generate and push the early warning information; as well as An updating module is used to update the model online.

Citation Information

Patent Citations

  • Systems and methods for real time classification and performance monitoring of batch processes

    CN102209958A

  • Multi-source multi-domain information entropy fusion and model self-optimization method for bearing fault diagnosis

    CN114048769A

  • Data anomaly detection method based on multi-scale residual classifier

    CN115733673A

  • Diagnostic tool to tool matching and full-trace drill-down analyasis methods for manufacturing equipment

    US20230236586A1

Cited By

  • Big data anomaly detection method based on time sequence multi-dimensional auto-encoder

    CN120105076A

  • Distributed surface well data processing method

    CN120234600A

  • A distributed well data processing method

    CN120234600B

  • Low-power-consumption signal detection method and system in metal card production process

    CN120492899A

  • Data mining method applied to hardware product control system

    CN121579853A