Deep Vision in Smart Manufacturing: MODERN Framework for Intelligent Quality Monitoring and Diagnosis

Background and Motivation

• Smart factories’ sensors and cameras continuously capture product dimensions, surface defects, and spatial patterns, requiring more responsive and accurate SPC control charts than traditional sampling-based designs.Traditional manufacturing samples only a fraction of products because human workers and analog machines cannot keep pace with high production volumes, whereas smart factories can monitor products continuously.Source: Background and Motivation · pp. 2–3
• The paper proposes a general framework combining artificial intelligence with statistical methods for image-based quality monitoring and fault diagnosis.The framework is motivated by the need to process high-volume camera images, inform control decisions promptly, and provide accurate post-signal diagnosis.Source: Background and Motivation · p. 3

1.2. Literature Review on Image-Based Quality Control

• Prior CNN-based industrial inspection studies determine whether products meet quality standards, but they primarily address defective-product identification rather than persistent process-shift monitoring.The paper’s main output is an image-based SPC control chart, not a product quality inspection technique.Source: §1.2 · pp. 3–4
• Existing image-monitoring methods impose restrictive assumptions on captured images, motivating a flexible approach that handles a wider variety of image types.The cited examples include deterministic gold-standard images, stochastic Markov Field images, and images lacking either property that existing methods cannot treat well.Source: §1.2 · p. 4
• Existing methods assume stationarity, locality, or gold-standard images, so they can fail on products whose images violate these conditions.The paper motivates a general CNN-based control-chart framework that handles varied images without restrictive assumptions.Source: §1.2 · p. 5

1.3. Results Overview and Contributions

• MODERN provides implementable image-based SPC for quality monitoring and post-signal diagnosis, using residual-network architecture and EWMA control charts to detect process shifts.The chart can detect small, persistent shifts and the diagnosis component locates defective regions in out-of-control images.Source: §1.3 · pp. 5–6
• MODERN-Diagnosis locates defective regions after an out-of-control signal and can be calibrated using a relatively small dataset of defective images.The framework is not tied to one neural-network choice and is intended to evolve with deep-learning technology.Source: §1.3 · p. 6
• Transfer monitoring combines an applicability test with augmented bootstrap control-limit determination, requiring fewer samples than retraining while achieving a faster convergence rate.The applicability test checks whether a pretrained network distinguishes in-control from out-of-control images and guards against framework misuse.Source: §1.3 · p. 6
• The framework establishes minimax-optimal convergence for defect-likelihood estimation and fault diagnosis, extending the relevant approximation result to CNNs.The errors approach zero with increasing sample size, and the CNN estimation rate cannot be improved by another method.Source: §1.3 · p. 7
• Mislabeling contributes separately to defect-likelihood estimation error, allowing managers to bound acceptable mislabeling relative to sample size.Deep-learning quality control becomes less effective when that bound is exceeded.Source: §1.3 · p. 7
• Compared with manual inspection, MODERN offers higher efficiency and repeatability by reducing fatigue and subjectivity in manufacturing quality control.Machines do not need rest, while human operators may make errors and differ in training or visual perception.Source: §1.3 · p. 7
• Because convergence worsens as image resolution increases, manufacturers should weigh image and labeling quality rather than automatically upgrading monitoring equipment.When mislabeling is absent or tolerable, not upgrading or even downgrading imaging equipment may be beneficial.Source: §1.3 · pp. 7–8
• When the sample size exceeds the asymptotic bound and the method becomes less effective, higher-resolution images are necessary, making equipment upgrades warranted.This condition qualifies the framework’s effectiveness and links improved image resolution to precise labeling.Source: §1.3 · p. 8
• Following an OC signal, MODERN-Diagnosis identifies the specific defective part, providing interpretable information about the root cause for quality managers.The diagnosis is intended to support interpretation of the MODERN-Chart signal and identify the faulty part that triggered it.Source: §1.3 · p. 8
• MODERN detects process shifts promptly and isolates faulty areas accurately across benchmark, simulated, and real industrial-image studies, outperforming MF in timely signaling and diagnosis.On DAGM, MODERN-Chart often signals at the exact change point and the faulty region estimator performs well with few OC images; on real data, MF may issue false alarms before shifts begin.Source: §1.3 · p. 8

1.4. Organization of the Paper

• The paper proceeds from the inception-residual deep CNN model and MODERN framework through theory, benchmark evaluation, numerical studies, conclusions, and appendix details.Sections 2–6 cover the model, framework, theoretical results, benchmark experiments, and summary; technical details are placed in the appendix.Source: §1.4 · pp. 8–9

2. Model Descriptions

• Deep CNNs learn representative image features from data, avoiding manual feature definitions that are time-consuming, domain-specific, and difficult to generalize.This approach addresses the challenge that image information depends on complex spatial patterns rather than isolated pixels.Source: §2 · p. 9

2.1. Basic Building Blocks

• The CNN uses convolutional, activation, pooling, and fully connected layers as its four transformation-layer types.Layer outputs feed subsequent layers, progressively extracting image attributes and producing interpretable outcomes.Source: §2.1 · p. 9

2.1.1. Convolutional Layer.

• Convolutional layers slide shared kernels across local neighborhoods to compute weighted sums, while stride, padding, kernel size, and filter count determine outputs.Locality captures nearby-pixel patterns, and translation-invariant parameter sharing reduces the number of parameters.Source: §2.1.1 · pp. 9–10

2.1.2. Activation Layer.

• The ReLU activation introduces nonlinear relationships and suppresses pixels lacking the kernel-defined feature, enabling detection of signal strengths and locations.Activation layers preserve input dimensions while transforming each input element through a nonlinear function.Source: §2.1.2 · pp. 10–11

2.1.3. Pooling Layer.

• Max and average pooling summarize neighborhoods and downsize feature maps while preserving detected features.Pooling uses neighborhood maxima or averages, with stride and neighborhood size controlling output dimensions.Source: §2.1.3 · p. 11

2.1.4. Fully Connected Layer.

• Fully connected layers flatten preceding outputs and map derived image features to target variables, using OC probability for quality monitoring and a five-dimensional faulty-region output for diagnosis.The OC probability is obtained through a logistic transformation, whereas the five-dimensional diagnosis output does not require it.Source: §2.1.4 · pp. 11–12

2.2. Inception Architecture with Residual Connections

• The framework motivates deep CNNs because shallow models may perform unsatisfactorily, whereas increasing depth enables greater approximation capability.Deep CNNs are described as having usually more than 100 convolutional layers.Source: §2.2 · p. 12
• The model adopts residual inception architecture because deep CNN training becomes harder with depth, while residual connections carry shallow features into deeper layers.Inception processes and concatenates features across multiple scales; residual connections add inputs to outputs before activation to preserve learned features.Source: §2.2 · pp. 12–13

2.3. The Network Architecture and Hyperparameters in Our Model

• MODERN-Net combines transformation layers with repeated InceptionRes modules in six main stages and contains about 23.4 million learnable parameters.The stages are Stem, InceptionRes-A, Reduction-A, InceptionRes-B, Reduction-B, and InceptionRes-C; InceptionRes-A, -B, and -C repeat five, ten, and five times, respectively.Source: §2.3 · p. 13
• After the main modules, MODERN-Net uses average pooling, a logistic fully connected layer, and dropout probability 0.6 before softmax to avoid overfitting.The complete MODERN-Net architecture and its hyperparameters are presented schematically in Figure 3.Source: §2.3 · p. 14, Figure 3 · p. 14

3.1. The Training of MODERN-Net

• MODERN-Net estimates the probability that an input product image contains defects by minimizing cross-entropy loss over labeled normal and defective images.Training uses stochastic gradient descent with randomly sampled mini-batches and a triangular cyclical learning-rate schedule; the resulting probability output constructs MODERN-Chart.Source: §3.1 · pp. 15–16

3.2. Quality Monitoring by MODERN-Chart

• MODERN-Chart recursively accumulates upward deviations of estimated defect probabilities from the in-control mean and signals when the accumulated statistic exceeds its control limit.The statistic resets to zero when new evidence is negative, helping the chart respond quickly to quality degradation.Source: §3.2 · pp. 16–17
• The Augmented Bootstrap algorithm selects MODERN-Chart’s control limit for a prespecified in-control average run length despite the chart statistic’s nonstationary distribution.It combines bootstrap resampling with six image transformations, expanding the in-control sample by a factor of six, and uses bisection search for fast convergence.Source: §3.2 · pp. 17–18

3.3. MODERN-Diagnosis: A Faulty Region Estimator

• After MODERN-Chart signals, the manufacturing process may stop and post-signal diagnosis begins to identify faulty regions for the ensuing repair process.The supplied passage introduces diagnosis as a critical and time-sensitive follow-up to monitoring.Source: §3.3 · p. 19
• MODERN-Diagnosis adapts the monitoring CNN to output five ellipse parameters that locate the defective region in an inspected image.The parameters describe the ellipse center, major and minor axes, and orientation; weighted L1 loss allows application-specific emphasis on these quantities.Source: §3.3 · p. 19
• Transfer learning initializes MODERN-Diagnosis from MODERN-Net features, enabling efficient training when defective-image data are much scarcer than non-defective images.Only the final layer is randomly initialized, while learned feature parameters are carried over from the monitoring network.Source: §3.3 · pp. 19–20

3.4. Transfer Monitoring: Extending the Applicability of MODERN Framework

• A pretrained network cannot be applied directly across manufacturing processes because image distributions and control limits may differ, even when classification remains applicable.The applicability test evaluates separation ability, whereas transfer monitoring recomputes the control limit for the new setting.Source: §3.4 · pp. 20–21
• Transfer monitoring uses a two-step procedure: hypothesis testing assesses whether a pretrained network applies to a new process, then new control limits are computed if applicability is supported.The procedure also provides a graphical applicability assessment and updates monitoring for the new setting.Source: §3.4 · pp. 20–21
• A few hundred new images can achieve commonly used ARL0 values such as 200, 500, or 1000, substantially extending MODERN to simulated and real-life settings.This requirement is much smaller than the tens of thousands of images typically needed to train a deep neural network from scratch.Source: §3.4 · p. 21

4. Asymptotic Properties

• The theoretical analysis uses techniques adapted from fully connected neural networks to establish minimax asymptotics for defect likelihood estimation and fault diagnosis.The results provide theoretical support for MODERN’s practical relevance and managerial implications.Source: §4 · p. 21

4.1. Theoretical Results in Estimation and Diagnostics

• MODERN-Net and MODERN-Diagnosis are asymptotically consistent, with estimation errors approaching 0 as training samples grow; MODERN-Net attains the minimax optimal rate up to a log factor.The results apply to Lipschitz targets, and the diagnosis estimator’s convergence is established under the stated theorem conditions.Source: §4.1 · p. 22, Theoretical implications · p. 23

Practical implications - part 1.

• Errors approach 0 as sample size grows, while convergence rates slow as image dimension d, equivalent to resolution, increases; the results indicate that low-resolution images can be beneficial.Source: Practical implications - part 1 · pp. 23–24
• When mislabeling is at or below N^-2/(2+d) log^3/2 N, asymptotic neural-network performance remains intact; above that rate, mislabeling degrades performance.This threshold quantifies when higher-resolution images and equipment upgrades become necessary for precise labeling.Source: Practical implications - part 1 · p. 24

4.2. Theoretical Results in Transfer Monitoring

• The AB algorithm achieves the typical convergence rate n^-1/2, much faster than N^-2/(2+d) log^3/2 N, the convergence rate in Theorem 1.Source: §4.2 · p. 25
• Theorem 4 provides the theoretical basis for the useful size-reduction property in transfer monitoring.Source: §4.2 · p. 25

5.1. The Impact of Image Resolution

• As the OC shift strengthens, low-resolution MODERN-Net detects the signal well, whereas high-resolution performance deteriorates, consistent with the theoretical findings.The finite-sample study simulates IC and OC images across three resolutions and summarizes classification accuracy in Table 1.Source: §5.1 · p. 26, Table 1 · p. 26

DAGM: A Benchmark Dataset.

• DAGM provides a challenging industrial benchmark because its synthetic textured-surface defects can be difficult for humans to discern and violate some Markov Field assumptions.The dataset contains 10 classes of textured surfaces with labeled elliptical defects and serves to evaluate quality monitoring and diagnosis.Source: DAGM: A Benchmark Dataset · p. 27

Numerical Setup and Results.

• MODERN’s control chart usually signals exactly at the change point, while MODERN-Diagnosis estimates defective regions reasonably well on the DAGM testing set.The study reports these monitoring and fault-isolation results using repeated simulations and average SDSC, with visual examples showing defective areas accurately highlighted despite imperfect ellipse overlap.Source: Table 2 · p. 28, Numerical Setup and Results · pp. 28–29

Numerical Setup and Results

• The numerical setup compares MODERN with Markov Field on simulated autoregressive images satisfying Markov Field’s locality and stationarity conditions, including noise-level and correlation defects.Because MODERN-Net was trained on DAGM, transfer monitoring and an applicability test precede application to the simulated AR images; the test indicates strong IC/OC discrimination for both defect types.Source: §5.3 · pp. 29–30

5.3. Comparison with Markov Field Method

• MODERN delivers signals more promptly than the MF approach and isolates faults reasonably well for both defect types in simulated monitoring and fault isolation.MODERN’s ARL1 standard error is zero, whereas MF almost never signals at the change point and has a large ARL1, indicating delayed signaling.Source: §5.3 · pp. 30–31

5.4. Application to Manufacturing of Electric Commutators

• For electric commutators, MODERN signals exactly at the change point and flags the region requiring attention, whereas Markov Field also produces a false signal.MODERN signals at t = 15; its estimated ellipse does not cover the entire defect but identifies the area requiring investigation, while Markov Field’s diagnosis is comparable.Source: §5.4 · pp. 31–32

5.5. Application to Leather Manufacturing

• In leather manufacturing, MODERN correctly detects a persistent process shift, while Markov Field produces false signals and intermittently returns to the in-control state.Both methods warn at t = 22, but only MODERN maintains the shifted-state indication; its EWMA mechanism accumulates out-of-control evidence, and both methods localize the defective region reasonably well.Source: §5.5 · p. 33

Contributions

• MODERN uses inception residual networks for fast abnormal-process signaling and effective defective-region identification after signals occur.The framework targets quality monitoring and fault diagnosis in smart manufacturing systems.Source: §6 · pp. 33–34
• Transfer monitoring supports insufficient image-data settings with small samples, alongside hypothesis testing to assess method applicability.Theoretical analysis establishes optimal convergence rates for defect-likelihood estimation and fault isolation.Source: §6 · p. 34

Managerial implication

• Image-label fidelity, rather than even free equipment upgrades, should guide monitoring and diagnosis decisions because upgrades may not improve quality control.This implication applies to both monitoring and diagnosis.Source: §6 · p. 34

Empirical evaluation

• Numerical comparisons show strong performance against a state-of-the-art approach, and real applications demonstrate MODERN in electric commutator and leather manufacturing.The authors plan randomized controlled experiments with industry partners to evaluate efficacy in real settings.Source: §6 · p. 34

Possible generalizations

• Beyond manufacturing, MODERN can support automated food inspection, smart-farming crop surveillance, and environmental monitoring of forestation, water resources, and urban landscapes.These applications use optical or image-sequence data to monitor relevant product, crop, or Earth-surface changes.Source: §6 · p. 34

Proof of Theorem 1

• The optimal discriminator satisfies D*=f* almost surely, as obtained from the convex population criterion by setting its first variation to zero.The proof identifies D* as the minimizer of the population objective and uses convexity to establish the equality.Source: Appendix A · p. 39
• Theorem 1 connects f* to minimization of the population cross-entropy loss and estimates it nonparametrically with ReLU neural networks.The estimator minimizes the corresponding sample cross-entropy over networks parameterized by depth D, width W, and bounded norm B.Source: Appendix A · p. 39

Neural-network estimation analysis

• The proof next bounds the L2(ν) estimation error between the target discriminator and its neural-network estimator.Here, ν denotes the marginal distribution used to define the nonparametric estimation norm.Source: Appendix A · p. 39
• The restricted population-loss landscape is shown to behave quadratically around the relevant solution, supporting the subsequent error analysis.The appendix introduces constants depending on B, including C1,...,C5, in this analysis.Source: Appendix A · p. 39

• The loss difference is decomposed into four terms involving bD_ϕ, b̄D_ϕ, and D∗.The decomposition separates successive differences between empirical and population losses and between candidate and target functions.Source: Proof of (6) · p. 40
• With Z = (X,Y) drawn from p(x,y), the proof establishes that b(D,Z) is 1-Lipschitz in D.Source: Proof of (6) · p. 40
• Independent copies of Zi and Rademacher variables σi(ϵi) are introduced to upper-bound the expected empirical-process term.The auxiliary variables are independent of Zi and each other for i = 1,...N.Source: Proof of (6) · p. 40
• The bound is expressed using the Gaussian complexity G(N) of ND,W,B.Source: Proof of (6) · p. 40

Proof of (9).

• The proof first applies Jensen’s inequality and distributional symmetry to rewrite the complexity expression involving b(D, eZ_i)−b(D,Z_i).The symmetry follows because the signed and unsigned differences are governed by the same law.Source: Proof of (9) · p. 41
• Lipschitz contraction bounds the Rademacher complexity using (8) and Theorem 12 of Bartlett and Mendelson (2002).The proof then relates Gaussian and Rademacher complexities using Lemma 4 of Bartlett and Mendelson (2002).Source: Proof of (9) · p. 41
• The argument next bounds the Gaussian complexity and obtains an expression involving N, D, W, and logarithmic factors.The displayed bound includes terms of the form (DW)^2 log(DW) log N and an exponential factor.Source: Proof of (9) · p. 41

Proof of (10).

• The neural-network complexity argument controls covering numbers through pseudo-dimension bounds for ReLU networks, using chaining and metric-entropy inequalities.The proof invokes Vershynin (2018), Anthony and Bartlett, and Bartlett et al. (2019) for these complexity steps.Source: Proof of (10) · pp. 41–42
• Lemma 3 bounds prediction error by empirical-process variation indexed by the network class plus approximation bias; combining Lemmas 3–5 yields the stated result.Lemma 5 gives a ReLU network with depth O(12D + C1,d) and width O(C2,dW).Source: Proof of (10) · pp. 42–43
• Theorem 2’s L1-loss convergence rate is slower than Theorem 1’s minimax-optimal rate because of technical difficulties from the L1 objective.The appendix states that adopting least-squares loss would recover the same optimal rate under the displayed proof strategy.Source: Proof of (10) · p. 44
• Theorem 3’s noisy-label analysis establishes an L2(ν) error bound of O(N^-2/(2+d) log^3/2 N) + O(δ).The proof defines a noisy logistic loss, relates it to the original loss, and controls the perturbation using boundedness of D.Source: Proof of (10) · p. 44
• The cited argument verifies that the average run length, ARL(ρ), equals M(F0), while the unsupported claims about a finite second moment and Gateaux differentials are removed.Source: Proof of (10) · p. 45
• The bootstrap EWMA result follows because the empirical in-control CDF converges uniformly to F0 at rate Op(1) after √nIC scaling by the Dvoretzky–Kiefer–Wolfowitz theorem.The proof applies the functional expansion around F0 to the empirical distribution generated by the bootstrap algorithm.Source: Proof of (10) · pp. 45–46

• DAGM background textures contain random patterns and may violate the Markov Field approach’s locality and stationarity assumptions, particularly in classes 4, 7, and 10.Source: §B.1 · p. 46
• DAGM images include surface defects that are difficult for human eyes to discern.Figure 8 presents examples of IC and OC images.Source: §B.1 · p. 46, Figure 8 · p. 47
• MODERN-Net training stops at epoch 150 for F(·;V) and epoch 275 for G(·;U), following stabilization from epochs 120 and 250, respectively.Figure 9 shows the training progressions used to select these stopping epochs.Source: §B.1 · p. 46, Figure 9 · p. 47

• The AR simulation generates IC images and two OC fault types satisfying the Markov Field conditions required by the MF method.Figure 10 presents representative images for the three categories.Source: §B.2 · p. 47, Figure 10 · p. 48
• A histogram visualization complements the p-value test by showing whether MODERN-Net reasonably separates IC and OC images.This visualization supports assessment of the network’s applicability to new image types.Source: §B.2 · p. 47
• Figure 11 shows that MODERN-Net distinguishes IC and OC AR images well despite not being trained on them, consistent with applicability-test statistics of 415.13 and 464.59 and effectively zero p-values.The test uses 500 IC images and 500 images of each OC fault type.Source: §B.2 · pp. 47–48, Figure 11 · p. 48
• MODERN-Net accurately locates the true faulty region for both type-1 and type-2 faults, while MF detects only the type-2 fault.In the type-1 case, MF detects no defective pixels.Source: §B.2 · p. 48, Figure 12 · p. 49

• Figure 13 shows one of the 14 IC images and the only OC image from electric commutator manufacturing.Source: Figure 13 · p. 49, §B.3 · p. 49
• The applicability test assesses whether the network trained on the benchmark dataset can be used in the considered manufacturing setting.Source: §B.3 · p. 49
• The trained network separated the OC image from the other images well, with an applicability-test statistic of 22.67 and p-value smaller than 10^-11.Figure 14 provides the direct visualization supporting the separation result.Source: §B.3 · p. 49

Benchmark design

• The study benchmarks MODERN-Net against AlexNet, a CNN architecture recognized as the winner of the 2012 ImageNet competition.The paper notes that other CNN architectures can also be used in the MODERN framework.Source: §B.4 · pp. 49–50

Architecture comparison

• AlexNet has 33.4 million parameters versus 23.4 million for MODERN-Net, despite being less deep.Both architectures use convolution layers, ReLU activations, and pooling operations.Source: §B.4 · p. 50, Table 4 · p. 51

Experimental setup

• AlexNet was trained with the same stochastic-gradient-descent method, triangular learning-rate schedule, and image transformations used for MODERN-Net.The fully connected layers were modified to process 512 × 512 images and output IC-OC probabilities.Source: §B.4 · p. 50

Comparison results

• MODERN-Net performed stably across the 10 image classes, whereas AlexNet’s performance varied substantially and was weak for OC classification.The comparison used the same training and validation DAGM datasets and evaluated both networks on the testing set.Source: §B.4 · p. 50, Table 5 · p. 51