Big data enabled smart factory whole-process online monitoring and early warning system

Abstract

The invention discloses a big data enabled smart factory full-process online monitoring and early warning system, and belongs to the technical field of industrial internet and intelligent manufacturing. According to the system, aiming at the problems of data fusion bottleneck, computing architecture limitation, intelligent analysis defects and the like of a traditional factory monitoring system, full-process monitoring and early warning are realized through a multi-source heterogeneous data fusion module, an edge-cloud collaborative architecture, an intelligent monitoring engine, a multi-mode early warning module and the like. The system adopts InfluxDB to realize millisecond-level data acquisition, utilizes a compression function to reduce data transmission quantity, combines a Spark framework to drive a three-dimensional digital twinning board, integrates an EWMA control chart and an LSTM neural network to carry out multi-modal early warning, realizes fault diagnosis and energy consumption anomaly detection through GCN, NMF and DTW, optimizes an early warning decision by means of a random forest, and realizes fault diagnosis and energy consumption anomaly detection. And the system is in butt joint with an industrial internet platform through a standardized API interface. The system can break through the bottleneck of data fusion, reduces the transmission and storage cost, improves the monitoring and early warning accuracy, and is used for various industrial scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of industrial internet and intelligent manufacturing technology, and particularly relates to a big data empowered intelligent factory whole-process online monitoring and early warning system. BACKGROUND

[0002] With the deepening of industrial digital transformation, the real-time, accuracy and intelligence of the whole-process online monitoring and early warning system of the intelligent factory are put forward with higher requirements. The traditional factory monitoring system faces three core challenges:

[0003] 1. Data fusion bottleneck: the factory environment contains PLC, MES, sensor and other multi-source heterogeneous data (≥11 types), the data type is complex and the collection frequency demand is high (the cycle needs to be ≤10 ms), and the existing system is difficult to realize millisecond-level time series data fusion;

[0004] 2. Computing architecture limitations: high cost of massive data transmission and storage, insufficient edge-cloud collaboration capabilities, and lack of effective data compression mechanisms to reduce network load;

[0005] 3. Intelligent analysis defects: traditional early warning relies on static thresholds (such as SPC control charts), lacks multi-modal analysis capabilities for equipment degradation trend prediction, energy consumption anomaly detection and other scenarios, and fault diagnosis relies on human experience, making it difficult to cover complex scenarios of 30+ equipment fault types.

[0006] The current technology has the following deficiencies:

[0007] 1. The data acquisition layer generally uses a single database, which cannot meet the millisecond-level write and query requirements of InfluxDB-level time series data;

[0008] 2. Edge computing only supports simple filtering, lacks efficient dimensionality reduction methods such as compression function Φ(x)=∥W T x+b∥ ∞ (reducing transmission by 90%, according to the document);

[0009] 3. Real-time calculation of dynamic indicators (such as ) relies on batch processing frameworks, making it difficult to drive real-time rendering of three-dimensional digital twin dashboards through Spark engines;

[0010] 4. The early warning module mostly uses single models (such as SPC or LSTM), without the fusion of exponential weighted moving average control charts (UCL t formula contains a dynamic decay factor λ) and LSTM cell state update mechanism (C t =f t ⊙C t-1 +i t ⊙tanh(·)), and the loss function does not introduce adaptive Huber loss To improve the prediction robustness;

[0011] 5. The fault diagnosis lacks deep integration of knowledge graph and graph convolution network (GCN), and cannot accurately predict the equipment failure trend through Implementing equipment failure reasoning;

[0012] 6. Energy consumption detection relies on threshold method, without applying quantitative analysis means such as non-negative matrix factorization (NMF) reconstruction error threshold (η>0.15 triggers early warning).

[0013] In addition, the existing system has not formed a full-process closed loop in key technical links such as equipment curve similarity matching (DTW algorithm is needed to minimize the cumulative distance of path ∑d(p k )≤∈), stress field simulation (based on physical engine to solve equation), early warning decision optimization (random forest selects split point with maximum ΔGini), etc. The industrial internet platform lacks standardized API interface constraints (such as data packet ∥d i ∥2≤Γ).

[0014] Therefore, in view of the above, the present application is proposed to improve the existing technology, and a big data enabled intelligent factory full-process online monitoring and early warning system is provided. SUMMARY

[0015] The technical problem to be solved by the present application is that the traditional factory monitoring system cannot cope with the millisecond-level (Δt≤10ms) real-time fusion and transmission pressure of multi-source heterogeneous data (≥11 types), and the high network load caused by the lack of edge-cloud collaborative mechanism; at the same time, the existing early warning method relies on a single model (such as static SPC threshold), which cannot realize accurate prediction of equipment degradation trend (such as dynamic update of LSTM cell state C t ), energy consumption anomaly detection (such as quantitative judgment of NMF reconstruction error η>0.15), and multi-fault mode reasoning (such as identification of 30+ fault types by GCN), resulting in low early warning accuracy and limited coverage scenarios, and it is urgent to solve the problems of real-time, intelligence and decision optimization through a big data enabled full-process closed loop system (from data acquisition to three-dimensional twin visualization).

[0016] The technical scheme adopted by the present application is: a big data enabled intelligent factory full-process online monitoring and early warning system, comprising:

[0017] Multi-source heterogeneous data fusion module: access at least 11 types of data sources such as PLC, MES and sensors through industrial internet gateway, realize millisecond-level data acquisition by using time series database (InfluxDB), and meet the acquisition period Δt≤10ms;

[0018] Edge-cloud collaborative architecture: edge node performs data preprocessing, through compression function Φ(x) = ||W T x + b||. ∞ Reduce 90% data transmission, cloud data center uses data lake storage;

[0019] Intelligent monitoring engine: based on Spark framework real-time computing dynamic indicators And drive three-dimensional digital twin board;

[0020] Multi-modal early warning module: integrate SPC threshold early warning and LSTM neural network prediction model.

[0021] As a further scheme of the application: the SPC threshold early warning adopts exponential weighted moving average (EWMA) control chart, and the control line calculation satisfies:

[0022]

[0023] Wherein μ0 is the baseline mean, σ is the standard deviation, L is the control coefficient, and λ is the weight attenuation factor.

[0024] As a further scheme of the application: the LSTM neural network prediction model realizes equipment degradation trend analysis through the following cell state update:

[0025] f t = σ(W f · [h t-1 , x t ] + b f )

[0026] C t = f t ⊙C t-1 + i t ⊙ tanh(W C · [h t-1 , x t ] + b C )

[0027] Wherein f t is the forget gate, i t is the input gate, W is the weight matrix, and h t is the hidden state.

[0028] As a further scheme of the application: the loss function of LSTM adopts adaptive Huber loss:

[0029] Other

[0030]

[0031] δ is a dynamic threshold parameter.

[0032] As a further scheme of the present application: adopt dynamic time warping (DTW) algorithm to match the similarity of equipment operation curve:

[0033]

[0034] Wherein dist is the Euclidean distance function, the cumulative path satisfies

[0035] As a further scheme of the present application: built-in industry knowledge graph realizes fault mode identification through graph convolution network (GCN):

[0036]

[0037] Wherein is the adjacency matrix, H (l) is the lth layer feature, supports 30+ equipment fault type reasoning.

[0038] As a further scheme of the present application: realize energy consumption anomaly detection through non-negative matrix factorization (NMF):

[0039]

[0040] When the reconstruction error , trigger the early warning.

[0041] As a further scheme of the present application: three-dimensional digital twin board adopts physics engine to realize real-time rendering of equipment stress distribution field:

[0042]

[0043] Wherein sigma is the Cauchy stress tensor, and u is the displacement field vector.

[0044] As a further scheme of the present application: optimize early warning decision through random forest algorithm:

[0045]

[0046] Select feature split point to make the impurity decrease Maximum.

[0047] As a further scheme of the present application: through API interface docking with Guolian video industrial internet platform, meet the transmission protocol:

[0048] Gamma is the data packet capacity constraint.

[0049] The beneficial effects of the present application are:

[0050] 1. Breakthrough multi-source heterogeneous data fusion bottleneck

[0051] Access to PLC, MES and sensors ≥11 types of data sources through industrial Internet of Things gateway, use time series database (InfluxDB) to realize millisecond level collection (period ≤10ms), solve the problem that traditional system cannot fuse high frequency heterogeneous data, provide complete data basis for whole process monitoring.

[0052] 2, greatly reduce the cost of data transmission and storage

[0053] Edge node performs data preprocessing, adopts compression function Φ(x) = ||W T x+b||. ∞ Realize 90% transmission compression, combined with cloud data lake storage, build edge-cloud collaborative architecture, significantly reduce network load and storage overhead.

[0054] 3, drive real-time dynamic monitoring and three-dimensional visualization

[0055] Based on Spark framework, real-time calculation of dynamic indicators Parallel three-dimensional digital twin board, realize physical engine level rendering of equipment running state (support stress field simulation ), improve monitoring response speed and visualization accuracy.

[0056] 4, multi-modal early warning improves accuracy and robustness

[0057] SPC threshold early warning: through exponential weighted moving average (EWMA) control chart dynamic adjustment control line (UCL t Containing attenuation factor λ), adapt to working condition fluctuation;

[0058] LSTM prediction model: combined with cell state update mechanism (C t = f t ⊙C t-1 +i t ⊙tanh(·)) and adaptive Huber loss function Enhance the robustness of equipment degradation trend prediction;

[0059] Double model fusion significantly reduces false alarm rate and covers complex fault scenarios.

[0060] 5, intelligent and knowledge-based fault diagnosis

[0061] Built-in industry knowledge graph, through graph convolution network (GCN) model Realize automatic reasoning of ≥30 types of equipment faults, reduce manual dependence and improve diagnosis coverage.

[0062] 6, quantitative detection of energy consumption and abnormal behavior

[0063] Use non-negative matrix factorization (NMF) model Detect energy consumption anomaly, trigger warning when reconstruction error η>0.15; Combine DTW algorithm D(i,j)=dist(x i ,y j )+min(·) to match device curve similarity, accurately identify abnormal operation mode.

[0064] 7、Optimize decision-making and industrial internet compatibility

[0065] Random forest algorithm selects feature split point with maximum Gini impurity reduction ΔGini, improves warning decision reliability;

[0066] Standardized API interface meets data packet constraint ∥d i ∥2≤Γ, realizes seamless connection with industrial internet platforms such as Guolian Video, and supports large-scale deployment. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 Flowchart of the present application, a big data enabled intelligent factory full-process online monitoring and early warning system. DETAILED DESCRIPTION

[0068] The present application will be further described below.

[0069] Please refer to Figure 1

[0070] Example 1: Monitoring of a machine shop (high-precision equipment health management)

[0071] Application scenario: automobile engine cylinder production line, which needs to monitor the vibration, temperature and tool wear state of numerical control machine tools in real time.

[0072] System configuration:

[0073] Edge layer: deploy industrial Internet gateway, access 11 types of data sources (PLC control signal, vibration sensor, infrared thermometer, etc.).

[0074] Cloud: use InfluxDB cluster (write speed ≥100,000 points / second), Spark computing node (32 cores / 128GB memory).

[0075] Warning module: LSTM prediction model (hidden layer 128 units) + SPC dynamic threshold (λ=0.2, L=2.7).

[0076] Implementation steps:

[0077] 1、Data acquisition: through the compression function Φ(x)=∥W T x+b∥ ∞ Compress vibration signal at edge (original data volume: 1.2GB / s→ compressed: 120MB / s).

[0078] 2. Dynamic indicator calculation: Real-time calculation of tool wear gradient (Q t Wear amount, Δt = 5 ms).

[0079] 3. Degradation prediction: LSTM model updates cell state C every 10 minutes t , combined with adaptive Huber loss (δ = 0.3) to predict remaining life.

[0080] 4. Early warning trigger: When the UCL of SPC t exceeds μ0+3σ or the LSTM prediction error exceeds the threshold continuously, trigger a three-level alarm.

[0081] Effect comparison:

[0082]

[0083]

[0084] Advantages: For high-value equipment, LSTM + SPC dual model reduces the failure false alarm rate from 15% to 2%.

[0085] Example 2: Energy efficiency monitoring of chemical plant (multi-node energy consumption optimization)

[0086] Application scenario: Ethylene cracking device, which needs to monitor 20+ parameters such as reaction kettle temperature, pressure, steam consumption, etc. simultaneously.

[0087] System configuration:

[0088] Edge layer: Custom compression matrix W ∈ R 50×5 (b = 0), transmission volume reduced by 92%.

[0089] Analysis module: NMF energy consumption model (k = 8) + DTW curve matching ( ∈ = 0.1).

[0090] Knowledge graph: GCN fault library (35 kinds of reactor fault modes).

[0091] Implementation steps:

[0092] 1. Abnormality detection: NMF decomposes energy consumption matrix V, when the reconstruction error locates the abnormal equipment.

[0093] 2. Pattern matching: DTW algorithm compares the current pressure curve with the historical fault library, and when the path cumulative distance ∑d(p k ) ≤ 0.08, it is determined as "catalyst coking".

[0094] 3. Three-dimensional simulation: Physical engine solves Visualize the stress hotspots of the reactor.

[0095] Effect comparison:

[0096] Indicator Threshold alarm system System of the present invention Abnormal positioning accuracy Equipment level Component level Number of false stops per month 9 times 1 time Steam consumption reduction — 8.7％

[0097] Innovation: NMF+DTW combination shortens the energy consumption anomaly detection response time from 30 minutes to 45 seconds.

[0098] Example 3: Electronic assembly line quality control (multi-source data fusion)

[0099] Application scenario: SMT patch production line, need to coordinate AOI detection, MES work order, temperature and humidity data.

[0100] System configuration:

[0101] Data lake: store InfluxDB time series data + relational MES data.

[0102] Decision module: Random Forest (100 trees, ΔGini>0.2 split).

[0103] Interface specification: API transmission meets ∥d i ∥2≤Γ=1.5MB.

[0104] Implementation steps:

[0105] 1. Knowledge graph diagnosis: GCN model update Identify "solder fake solder" faults.

[0106] 2. Dynamic threshold adjustment: EWMA control chart updates μ0(λ=0.15) with batch changes.

[0107] 3. Decision optimization: Random Forest automatically adjusts the warning strategy according to feature importance (such as solder paste thickness, reflow soldering temperature).

[0108] Effect comparison:

[0109] Indicator Single model system System of the present invention Defect detection rate 88％ 99.2％ Fault classification error 5 types of confusion Zero confusion Platform docking cycle 3 weeks 72 hours

[0110] Technical breakthrough: GCN+Random Forest improves fault classification accuracy from 91% to 98.5%.

[0111] Example comparison analysis

[0112] Dimension Example 1 (mechanical) Example 2 (chemical) Example 3 (electronic) Core module LSTM+SPC NMF+DTW GCN+random forest Delay sensitivity Extremely high (Δt≤5ms) Medium (1 minute level) Low (5 minute level) Data processing focus Time series signal compression Matrix decomposition Heterogeneous data fusion Fault coverage Equipment level (12 types) System level (35 types) Process level (28 types) ROI improvement Device maintenance cost ↓ 37% Energy consumption cost ↓ 22% Quality inspection cost ↓ 41%

[0113] Industrial Internet docking implementation

[0114] All examples are connected with the Guolian video platform through standardized API interface:

[0115] Φ={(di ,t i )|d i ∈R n ,||d i ||2≤Γ}.

[0116] Docking process:

[0117] 1. Edge node encapsulates data packet d i (Satisfy ∥d i ∥2≤1.5MB).

[0118] 2. Cloud access Normalize device topology data.

[0119] 3. Dynamically adjust the compression matrix W to adapt to the new device data characteristics.

[0120] Verification results: Deployment in 15 factories shows that the system supports cross-scenario migration from discrete manufacturing to process industries, with an average deployment cycle of ≤7 days.

[0121] The technical solution of the present invention has verified its universality and technical superiority in complex industrial scenarios through the above-mentioned embodiments. In practical applications, the combination of functional modules can be selected according to needs to achieve the optimal balance between monitoring accuracy and cost.

[0122] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A big data-enabled smart factory full-process online monitoring and early warning system, characterized by: include: Multi-source heterogeneous data fusion module: Accesses at least 11 data sources, including PLCs, MES, and sensors, through the Industrial Internet of Things gateway. It uses a time series database (InfluxDB) to achieve millisecond-level data collection, ensuring a collection cycle of Δt ≤ 10ms. Edge-cloud collaborative architecture: Edge nodes perform data preprocessing and pass compression function Φ(x)=∥W T x+b∥ ∞ Reduce data transmission volume by 90% by using data lake storage in cloud data centers; Intelligent monitoring engine: Real-time calculation of dynamic indicators based on the Spark framework and drives the 3D digital twin dashboard; Multimodal warning module: integrates SPC threshold warning and LSTM neural network prediction model.

2. The big data-enabled smart factory full-process online monitoring and early warning system according to claim 1 is characterized in that: The SPC threshold warning adopts the exponentially weighted moving average (EWMA) control chart, and its control line calculation satisfies: Where μ0 is the benchmark mean, σ is the standard deviation, L is the control coefficient, and λ is the weight attenuation factor.

3. The big data-enabled smart factory full-process online monitoring and early warning system according to claim 1 is characterized in that: The LSTM neural network prediction model implements device degradation trend analysis through the following cell state updates: f t =σ(W f ·[h t-1 ,x t ]+b f ) C t =f t ⊙C t-1 +i t ⊙tanh(W C ·[h t-1 ,x t ]+b C ) where f t For the forget gate, i t is the input gate, W is the weight matrix, h t In hidden state.

4. The big data-enabled smart factory full-process online monitoring and early warning system according to claim 1 is characterized in that: The loss function of LSTM uses adaptive Huber loss: other δ is the dynamic threshold parameter.

5. The big data-enabled smart factory full-process online monitoring and early warning system according to claim 1 is characterized in that: Use the Dynamic Time Warping (DTW) algorithm to match the similarity of equipment operation curves: Where dist is the Euclidean distance function, the cumulative path satisfies 6. The big data-enabled smart factory full-process online monitoring and early warning system according to claim 1 is characterized in that: Built-in industry knowledge graph uses graph convolutional networks (GCN) to achieve fault pattern recognition: in is the adjacency matrix, H (l) It is the first layer feature and supports reasoning about 30+ device fault types.

7. The big data-enabled smart factory full-process online monitoring and early warning system according to claim 1 is characterized in that: Energy consumption anomaly detection through non-negative matrix factorization (NMF): When the reconstruction error When the warning is triggered.

8. The big data-enabled smart factory full-process online monitoring and early warning system according to claim 1 is characterized in that: The 3D digital twin dashboard uses a physics engine to render the equipment stress distribution field in real time: where σ is the Cauchy stress tensor and u is the displacement field vector.

9. The big data-enabled smart factory full-process online monitoring and early warning system according to claim 1 is characterized in that: Optimizing early warning decisions through random forest algorithms: Select feature splitting points to reduce impurity maximize.

10. The big data-enabled smart factory full-process online monitoring and early warning system according to claim 1 is characterized in that: Connect with Guolian Video's industrial Internet platform through API interface to meet the transmission protocol: Γ is the data packet capacity constraint.

Images