Abnormality monitoring method and system for aluminum plate cutting equipment

Abstract

The invention discloses an aluminum plate cutting equipment abnormity monitoring method which comprises the following steps: S1, acquiring cutting pressure time sequence data, a section texture image and equipment current / voltage signals in real time, and calling a pre-stored aluminum plate material characteristic library, cutting pressure time sequence data Pt output by a pressure sensor and an aluminum plate section RGB image I captured by an industrial camera; the material characteristic library comprises hardness H, thickness d and surface friction coefficient; and S2, constructing a knowledge graph in the graph database. Through a dual-channel collaborative sensing mechanism, the inherent defects of traditional single-path monitoring are thoroughly overcome, illumination self-adaptive processing is adopted in a visual channel, section texture features can still be stably captured under extreme working conditions such as strong light and dim light, the defect that a conventional visual system is seriously interfered by the environment is overcome, and the visual sensing accuracy is improved. And the pressure channel analyzes the dynamic cutting process by using a time sequence deep network, and the sudden abnormal impact is accurately identified.

Description

Technical Field

[0001] This invention relates to the field of aluminum plate cutting monitoring technology, specifically to a method and system for abnormal monitoring of aluminum plate cutting equipment. Background Technology

[0002] Aluminum plate cutting equipment plays a crucial role in modern industrial production, and its performance stability and cutting precision directly affect product quality and production efficiency. However, in actual production, the performance of aluminum plate cutting equipment is often affected by various factors, such as material characteristics, cutting parameters, equipment aging, and improper operation, all of which can lead to equipment malfunctions.

[0003] Traditional anomaly detection methods suffer from three major technical bottlenecks:

[0004] Single perception dimension: Existing technologies (such as CN112276298A) mainly rely on pressure sensors or visual inspection as a single path: pressure monitoring cannot identify texture abnormalities caused by micro-wear of the tool, and the vision system is severely affected by light interference (the false alarm rate is as high as 30%-40% in strong light environment).

[0005] Rigid decision-making mechanism: The mainstream solution adopts a fixed threshold for judgment: the threshold setting depends on the engineer's experience and cannot adapt to changes in material properties. Research shows that when the aluminum plate thickness increases from 3mm to 12mm, the false alarm rate of the static threshold solution increases.

[0006] Closed systems lack evolution: Industry-standard systems (such as Siemens CutMonitor) have significant flaws: they cannot integrate process knowledge bases, mistakenly identify abnormal material impurities as tool failures (a case in a car company led to a 22-hour downtime), lack online learning mechanisms, require manual recalibration of parameters every 3 months, and maintenance costs account for 35% of the total system cost. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for abnormal monitoring of aluminum plate cutting equipment, so as to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method and system for abnormal monitoring of aluminum plate cutting equipment, comprising the following steps:

[0009] S1. Real-time acquisition of cutting pressure timing data, cut surface texture image, and equipment current / voltage signal; and retrieval of pre-stored aluminum plate material property library, cutting pressure timing data Pt output by pressure sensor, and RGB image I of aluminum plate cut surface captured by industrial camera. The material property library includes hardness H, thickness d, and surface friction coefficient. ;

[0010] S2. Constructing knowledge graphs from graph databases, including:

[0011] Entities: {Aluminum plate type, cutting tool, equipment parameters, fault type};

[0012] Relationship: The hardness range and material properties are derived from the aluminum plate type; the applicable aluminum plate thickness is derived from the cutting tool; and the type of abnormality is derived from the equipment parameters.

[0013] S3. For the pressure time series channel: Use a temporal convolutional network (TCN) to process the pressure sequence Pt, output the anomaly score Sp, and calculate the variance. ;

[0014] For the visual channel: convert I to HSV space and extract the luminance component of the V channel;

[0015] Gamma correction is applied to the V channel: ;

[0016] Calculate HOG feature descriptors for the corrected image;

[0017] Input the MobileNetV3 network and output the visual anomaly score Sv;

[0018] Calculate the variance of HOG eigenvectors v;

[0019] S4. Based on the material property vector M=[H,d, ] and feature dispersion D=[ , v], calculate the dynamic threshold T using a piecewise function:

[0020]

[0021] in and As the baseline threshold, Here, W is the sigmoid function, and W is the weight matrix. For bias terms;

[0022] S5. Atlas query threshold adapts to thickness:

[0023] When d≤5mm: threshold=0.7;

[0024] When d > 5 mm: threshold = 0.8 + 0.02 × (d - 5);

[0025] Perform a graph query:

[0026] Input: Sp,threshold;

[0027] Output: Root cause set F = {f | Association confidence > threshold};

[0028] Calculate the confidence level of the spectrum:

[0029]

[0030] S6. Calculate the final anomaly score:

[0031]

[0032] in, , ;

[0033] S7, when >At time T, execute:

[0034] Differentiated shutdown strategy;

[0035] Weight update:

[0036]

[0037] ∈[0.01,0.1];

[0038] Confidence update:

[0039]

[0040] n represents the number of historical verifications;

[0041] The map update includes:

[0042] Added anomaly type condition: No anomalies were identified for ≥3 consecutive times.

[0043] Confidence decay: Unverified relationships decay by 10% per month.

[0044] In step S4:

[0045] when .

[0046] The construction of the material property library in step S1 includes: pre-storing a hardness-thickness-friction coefficient mapping table for ≥50 aluminum plate models, supporting synchronous updates with the MES system via the OPCUA protocol. The update formula for the material property library is:

[0047]

[0048] in The adaptive learning rate has a value range of [0.01, 0.1].

[0049] The abnormal entity types are classified into three levels: equipment failure - tool abnormality - {chipping, wear, fracture}; the process document is analyzed and disambiguated using the BERT model.

[0050] An abnormality monitoring system for aluminum plate cutting equipment, used to implement the monitoring method, includes:

[0051] Multi-source sensing module: including piezoelectric pressure sensor, 5-megapixel industrial camera and vibration sensor;

[0052] Edge computing module: includes an embedded hardware platform and deploys a TCN network;

[0053] Knowledge graph engine: Neo4j graph database, real-time Cypher query interface. The knowledge graph engine includes a process knowledge extractor, generates triples by parsing equipment manuals through OCR, and supports multi-tenant isolated storage.

[0054] Cloud-based optimization module: includes a distributed material property database and a graph self-learning engine.

[0055] The knowledge graph engine includes: a process knowledge extractor and a multi-tenant isolation module.

[0056] Compared with the prior art, the beneficial effects of the present invention are:

[0057] I. Comprehensive Breakthrough in Testing Capacity Bottlenecks

[0058] This invention completely solves the inherent defects of traditional single-path monitoring through a dual-channel collaborative sensing mechanism: the visual channel adopts illumination-adaptive processing, which can stably capture the texture features of the cut surface even under extreme conditions such as strong light and dim light, overcoming the drawback of conventional vision systems being severely affected by environmental interference. The pressure channel uses a temporal depth network to analyze the dynamic cutting process and accurately identify sudden abnormal impacts. The dual-channel data is fused for decision-making based on confidence, which not only avoids misjudgment of a single signal, but also enables comprehensive diagnosis of complex anomalies such as tool wear and material impurities, significantly improving the integrity of the detection.

[0059] II. Dynamically adapting to complex industrial scenarios

[0060] The parameter self-adjustment system of this invention enables the system to adapt to environmental conditions: Material characteristic response: The threshold calculation mode is automatically switched according to the hardness of the aluminum plate, and a more sensitive judgment strategy is adopted for high-hardness plates, effectively preventing false negatives. Thickness adaptation: The knowledge graph query threshold is dynamically adjusted according to the thickness of the aluminum plate to ensure accurate monitoring of differences between thin and thick plates. Fault tolerance mechanism: When the sensor malfunctions, a backup signal source is automatically activated to maintain continuous system operation and significantly reduce unplanned downtime.

[0061] III. A Continuously Evolving Intelligent Core

[0062] The system achieves self-iterative upgrades through a dual-loop optimization mechanism: Parameter loop: Based on real-time diagnostic errors, the feature weight matrix is ​​adjusted in reverse, enabling the model to continuously adapt to dynamic factors such as equipment aging and process changes. Knowledge loop: The material property database is automatically updated, and the knowledge graph relationships are optimized, adding unidentified anomaly types and periodically eliminating low-confidence knowledge nodes. This evolutionary capability significantly reduces the frequency of manual maintenance, making the system increasingly accurate with use. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] Please see Figure 1 The present invention provides a technical solution: a method for abnormal monitoring of aluminum plate cutting equipment, comprising the following steps:

[0066] S1. Real-time acquisition of cutting pressure timing data, cut surface texture image, and equipment current / voltage signal; retrieval of pre-stored aluminum plate material property library, cutting pressure timing data Pt output by pressure sensor (sampling rate 1kHz), and RGB image I of aluminum plate cut surface captured by industrial camera; the material property library includes hardness H, thickness d, and surface friction coefficient. ;

[0067] S2. Constructing knowledge graphs from graph databases, including:

[0068] Entities: {Aluminum plate type, cutting tool, equipment parameters, fault type};

[0069] Relationship: The hardness range and material properties are derived from the aluminum plate type; the applicable aluminum plate thickness is derived from the cutting tool; and the type of abnormality is derived from the equipment parameters.

[0070] S3. For the pressure time series channel: Use a temporal convolutional network (TCN) to process the pressure sequence Pt, output the anomaly score Sp, and calculate the variance. ;

[0071] For the visual channel: convert I to HSV space and extract the luminance component of the V channel;

[0072] The gamma correction factor γ is dynamically selected based on the ambient light intensity L:

[0073] When L < 100 Lux: γ = 1.2;

[0074] When 100 Lux ≤ L ≤ 1000 Lux: γ = 0.6;

[0075] When L > 1000 Lux: γ = 0.4;

[0076] Gamma correction is applied to the V channel: ;

[0077] Calculate HOG feature descriptors (8×8 pixels per cell unit) for the corrected image.

[0078] Input the MobileNetV3 network and output the visual anomaly score Sv;

[0079] Calculate the variance of HOG eigenvectors v;

[0080] S4. Based on the material property vector M=[H,d, ] and feature dispersion D=[ , v], calculate the dynamic threshold T using a piecewise function:

[0081]

[0082] in and As the baseline threshold, Here, W is the sigmoid function, and W is the weight matrix. For bias terms;

[0083] when ;

[0084] S5. Execute the graph query to output the root cause type set F, and calculate the graph confidence. ;

[0085] Atlas query thresholds adapt to thickness:

[0086] When d≤5mm: threshold=0.7;

[0087] When d > 5 mm: threshold = 0.8 + 0.02 × (d - 5);

[0088] S6. Calculate the final anomaly score:

[0089]

[0090] in, , ;

[0091] S7, when At time T, a differentiated shutdown strategy is executed based on F, and the graph is updated, as follows:

[0092] Weight update:

[0093]

[0094] ∈[0.01,0.1];

[0095] Confidence update:

[0096]

[0097] n represents the number of historical verifications;

[0098] When a sensor fails, the fault tolerance mechanism is activated:

[0099] Pressure sensor malfunction: Sp is replaced with the average of the last 10 cuts. Take the historical maximum value;

[0100] Industrial camera malfunction: Sv is calculated from the spectral entropy value of the current signal, Sv=1- ;

[0101] The map update includes:

[0102] Added anomaly type condition: No anomalies were identified for ≥3 consecutive times.

[0103] Confidence decay: Unverified relationships decay by 10% per month.

[0104] The construction of the material property library in step S1 includes: pre-storing a hardness-thickness-friction coefficient mapping table for ≥50 aluminum plate models, supporting synchronous updates with the MES system via the OPCUA protocol. The update formula for the material property library is:

[0105]

[0106] in The adaptive learning rate has a value range of [0.01, 0.1].

[0107] The abnormal entity types are classified into three levels: equipment failure - tool abnormality - {chipping, wear, fracture}; the process document is analyzed and disambiguated using the BERT model.

[0108] The incremental learning employs a sliding window mechanism: it stores the most recent N=1000 sets of outlier data samples, and the weight matrix update formula is:

[0109]

[0110]

[0111] When the false alarm rate within the window exceeds 5%, the TCN network will be automatically retrained.

[0112] An abnormality monitoring system for aluminum plate cutting equipment includes:

[0113] Multi-source sensing module: including piezoelectric pressure sensor, 5-megapixel industrial camera and vibration sensor;

[0114] Edge computing module: includes an embedded hardware platform and deploys a TCN network;

[0115] Knowledge graph engine: Neo4j graph database, real-time Cypher query interface. The knowledge graph engine includes a process knowledge extractor, generates triples by parsing equipment manuals through OCR, and supports multi-tenant isolated storage.

[0116] Cloud-based optimization module: includes a distributed material property database and a graph self-learning engine.

[0117] The edge computing module includes a fault diagnosis unit: when the vibration sensor detects a characteristic frequency f that satisfies:

[0118]

[0119] in The tool's fundamental frequency is used to determine tool wear and trigger an early warning, while the confidence weight α=0.2 is dynamically adjusted.

[0120] The knowledge graph engine includes: a process knowledge extractor and a multi-tenant isolation module;

[0121] The response rules for differentiated shutdown:

[0122] The machine should stop immediately when "Broken Blade" ∈ F;

[0123] When "wear" ∈ F, reduce the speed to 70%;

[0124] Standard shutdown procedures apply to other situations.

[0125] The workflow is as follows:

[0126] Multi-source sensor module installation:

[0127] A Kistler 9129AA piezoelectric pressure sensor was installed on the spindle of the cutting machine tool, with a sampling rate set to 1 kHz and a range of 0-5000 N.

[0128] A Baslerac A1920-150um industrial camera (5 megapixels) with a frame rate of 30fps is deployed 300mm from the cut surface of the aluminum plate, equipped with a ring light.

[0129] A HIOKICT7663 current sensor with an accuracy of ±0.5% is installed in the power supply circuit of the equipment.

[0130] A TSL2561 ambient light sensor with a range of 0-2000 Lux was fixedly installed next to the camera.

[0131] Material property library construction:

[0132] It pre-stores parameters for 52 common aluminum plate models, including hardness (HB), thickness (mm), and surface friction coefficient μ.

[0133] Taking 6061-T6 aluminum plate as an example: hardness 95HB, thickness 8.0mm, coefficient of friction 0.32.

[0134] Connect to the factory's MES system via the OPC-UA protocol;

[0135] Knowledge graph initialization:

[0136] Building entity relationship networks using the Neo4j graph database:

[0137] Create aluminum plate model node: 6061-T6 (hardness range 90-100HB).

[0138] Create cutting tool node: TC-15 (applicable thickness 5-10mm);

[0139] Create an exception type node: Collapse Blade (Level "Severe");

[0140] Establishing relationships:

[0141] (6061-T6)-[Applicable Cutting Tools]->(TC-15) Confidence level 0.9;

[0142] (6062-(TC-15)-[Causes Abnormality]->(Broken Blade) Confidence 0.85;

[0143] The device manual was automatically generated using OCR to analyze the equipment and generate 2000+ triples.

[0144] Visual channel processing:

[0145] Real-time detection of ambient light intensity L:

[0146] When L < 100 Lux, set the gamma coefficient γ = 1.2 (to enhance low-light contrast).

[0147] When 100≤L≤1000Lux, γ=0.6 (reference value);

[0148] When L>1000Lux, γ=0.4 (to suppress overexposure of strong light);

[0149] Perform gamma correction:

[0150] Convert the RGB image to HSV space, extract the V channel luminance component, and calculate the correction value;

[0151] HOG feature descriptors are computed on 8×8 pixel cell units;

[0152] Input a lightweight MobileNetV3 network and output a visual anomaly score Sv;

[0153] Pressure channel processing:

[0154] Processing the stress sequence Pt using a Temporal Convolutional Network (TCN):

[0155] Network structure: 4 layers of dilated convolutions (dilation factor [1,2,4,8]);

[0156] Input: 1×1000 data points (1-second timing sequence);

[0157] Output: The abnormal score Sp of the sigmoid activation (range 0-1);

[0158] Calculate the variance of the pressure series As a feature of dispersion;

[0159] Taking the cutting of 6061-T6 aluminum plate as an example (H=95HB>80HB):

[0160] Material property vector M = [95, 8.0, 0.32];

[0161] Characteristic dispersion D = [0.15, 0.08] (from pressure variance and HOG variance);

[0162] The weight matrix W2 = [0.3, -0.2, 0.4]ᵀ, and the bias term b = (0.5 - 0.1) / 2 = 0.2;

[0163] calculate: =0.3×95-0.2×8+0.4×0.32+0.2=28.93;

[0164] Sigmoid output: If (28.93)≈1.0>0.99, then the extreme condition is triggered;

[0165] With a baseline threshold T2 = 0.8, the final threshold T = 1.05 × 0.8 = 0.84;

[0166] When a camera malfunction is detected (heartbeat signal loss):

[0167] The signal is acquired from the current sensor, and the spectral entropy value Entropy is calculated.

[0168] Convert visual abnormality scores;

[0169] The abnormal score Sp of the pressure channel is the average of the last 10 cuts.

[0170] Feature Dispersion Take the historical maximum value.

[0171] Example 1: Detection of chipped blade anomalies (situation of sudden light change)

[0172] Operating conditions:

[0173] Cutting a 5052-H32 aluminum plate (4mm thick) with ambient light suddenly increasing from 800Lux to 1200Lux;

[0174] Processing procedure:

[0175] Light sensor triggers γ coefficient switching: 0.6-0.4;

[0176] HOG features detected a 37% increase in the texture breakage index;

[0177] The pressure channel captures the peak impact value (Sp=0.91).

[0178] The knowledge graph query outputs the root cause F={"collapsed blade"}, with a confidence level Ck=0.88;

[0179] Final outlier score =0.82>Dynamic threshold T=0.78;

[0180] Perform the following actions:

[0181] Stop the machine immediately to prevent complete damage to the cutting tools;

[0182] Economic benefits:

[0183] Reduce equipment maintenance losses by ¥15,000;

[0184] Example 2: Fault tolerance mechanism verification (sensor failure);

[0185] Simulated fault:

[0186] Manually disconnect the pressure sensor communication cable;

[0187] System response:

[0188] Automatically enable fault tolerance mode:

[0189] Sp = the average of the last 10 times (0.24);

[0190] =Historical maximum value (0.18);

[0191] Detection of abnormal harmonics using current spectrum entropy (Sv=0.76).

[0192] The graph query outputs F={"wear"}, with a confidence level Ck=0.75;

[0193] Implement a 70% speed reduction operation strategy.

[0194] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

Claims

1. A method for monitoring abnormalities in aluminum plate cutting equipment, characterized in that, Includes the following steps: S1. Real-time acquisition of cutting pressure timing data, cut surface texture image, and equipment current / voltage signal; and retrieval of pre-stored aluminum plate material property library, cutting pressure timing data Pt output by pressure sensor, and RGB image I of aluminum plate cut surface captured by industrial camera. The material property library includes hardness H, thickness d, and surface friction coefficient. ; S2. Construct a knowledge graph in the graph database, which includes the entity relationships between aluminum plates and cutting tools; S3. For the pressure time series channel: Use a temporal convolutional network (TCN) to process the pressure sequence Pt, output the anomaly score Sp, and calculate the variance. ; For the visual channel: convert I to HSV space and extract the luminance component of the V channel; S4. Based on the material property vector M=[H,d, ] and feature dispersion D=[ , v], calculate the dynamic threshold T using a piecewise function: S5. The map query threshold is adaptively adjusted according to thickness. The map query is executed, and the map confidence is calculated. S6. Calculate the final anomaly score: in, , ; S7, when At time T, a differentiated shutdown strategy is executed, including weight updates and confidence level updates.

2. The method according to claim 1, characterized in that: The atlas is configured to be updated, including: New anomaly type condition: No anomalies were identified for ≥3 consecutive times; Confidence decay: Unverified relationships decay by 10% per month.

3. The method according to claim 1, characterized in that: In step S4: and As the baseline threshold, Here, W is the sigmoid function, and W is the weight matrix. For bias terms, when .

4. The method according to claim 1, characterized in that: The construction of the material property library in step S1 includes: pre-storing a hardness-thickness-friction coefficient mapping table for ≥50 aluminum plate models, supporting synchronous updates with the MES system via the OPCUA protocol. The update formula for the material property library is: in The adaptive learning rate has a value range of [0.01, 0.1].

5. The method according to claim 1, characterized in that: The abnormal entity types adopt a three-level classification structure: equipment failure - tool abnormality - specific abnormal form, where the specific abnormal forms include chipping, wear, and breakage; the entity disambiguation of the knowledge graph is achieved by parsing the process documents using the BERT model.

6. The method according to claim 1, characterized in that: The extraction of the V channel luminance component specifically includes the following steps: Gamma correction is applied to the V channel: ; Calculate HOG feature descriptors for the corrected image; Input the MobileNetV3 network and output the visual anomaly score Sv; Calculate the variance of HOG eigenvectors v.

7. The method according to claim 1, characterized in that: The atlas query threshold is adaptive based on thickness: When d≤5mm: threshold=0.7; When d > 5 mm: threshold = 0.8 + 0.02 × (d - 5) Perform a graph query: Input: Sp,threshold; Output: Root cause set F = {f | Association confidence > threshold} Calculate the confidence level of the spectrum: 。 8. The method according to claim 1, characterized in that: The weight update: ∈[0.01,0.1]； The confidence update: n represents the number of historical verifications.

9. An abnormality monitoring system for aluminum plate cutting equipment, used to implement the monitoring method according to any one of claims 1-8, characterized in that: include: Multi-source sensing module: including piezoelectric pressure sensor, 5-megapixel industrial camera and vibration sensor; Edge computing module: includes an embedded hardware platform and deploys a TCN network; Knowledge graph engine: Neo4j graph database, real-time Cypher query interface. The knowledge graph engine includes a process knowledge extractor, generates triples by parsing equipment manuals through OCR, and supports multi-tenant isolated storage. Cloud-based optimization module: includes a distributed material property database and a graph self-learning engine.

Images