Dual-Granularity Guidance Network for Fault Diagnosis

• The paper introduces DGGN as an incremental fault diagnosis system that separates fine-grained discrimination from coarse-grained stability to address forgetting and overfitting.
• It employs a multi-semantic cross-attention mechanism to fuse class-specific and class-agnostic features, enhancing representation learning in few-shot settings.
• The architecture integrates a MOIA-enhanced, discriminative branch with a self-supervised, stable branch, supported by boundary-aware exemplar replay and a balanced random forest classifier.

Searching arXiv for the cited DGGN paper and closely related dual-granularity guidance work. Found the core DGGN paper and several adjacent dual-granularity methods for context. Proceeding with the article grounded in these sources. Dual-Granularity Guidance Network (DGGN) is a framework for Few-shot Class-incremental Fault Diagnosis (FSC-FD) that explicitly separates representation learning into a fine-grained representation stream for class-specific discrimination and a coarse-grained representation stream for class-agnostic knowledge preservation, then fuses the two through a multi-semantic cross-attention mechanism so that stable coarse-grained knowledge guides the learning of fine-grained features (Yang et al., 16 Aug 2025). In the paper that introduces the term, DGGN is not a standalone guidance block but a full incremental diagnosis system comprising dual-stream feature learning, cross-attentive semantic fusion, exemplar replay, and a decoupled classifier. Its stated purpose is to address the joint difficulty of catastrophic forgetting of previously learned faults and overfitting to scarce new-class samples in industrial diagnosis settings (Yang et al., 16 Aug 2025).

1. Problem formulation and conceptual basis

DGGN is formulated for sequential class-incremental learning over sessions

$$\mathcal{T} = \left[(C^1, D^1), (C^2, D^2), \ldots, (C^n, D^n)\right],$$

where $C^t$ is the set of classes introduced in session $t$ and $D^t$ is the corresponding training set. For the base session,

$$D^1 = \left\{ (x_1, y_1), (x_2, y_2), \ldots, (x_i, y_i),\; y_i \in C^1 \right\},$$

while for incremental sessions $t>1$,

$$D^t = \left\{ (x_i, y_i)_{i=1}^N,\; y_i \in C^t \right\}.$$

The class sets are disjoint across sessions,

$$\forall t_1 \neq t_2,\quad C^{t_1} \cap C^{t_2} = \emptyset,$$

and after session $n$ the model must classify over

$$C^{(0:n)} = \left\{ C^1 \cup C^2 \cup \cdots \cup C^n \right\}$$

(Yang et al., 16 Aug 2025).

The motivation is specific to industrial fault diagnosis. The paper states that fault samples are naturally scarce because systems mostly operate normally, and fault classes evolve over time because of aging, operating condition changes, or new failure modes. In this setting, a single-stream model is described as prone to representation entanglement and feature conflict: it may overfit to the few new samples while simultaneously distorting representations needed for old classes (Yang et al., 16 Aug 2025).

The central conceptual move in DGGN is to distinguish two representation types that prior FSC-FD methods tend to mix in one feature space. The class-specific / fine-grained stream is intended to remain plastic and discriminative for new faults. The class-agnostic / coarse-grained stream is intended to remain stable across sessions and capture general semantics shared across fault types. This division is the basis for the paper’s claim that forgetting and overfitting should be addressed jointly through representation design rather than solely through replay or distillation (Yang et al., 16 Aug 2025).

2. System architecture and dual-stream organization

The architecture is described at a high level as consisting of sample replay, feature expansion, and classifier. End to end, DGGN processes input fault signals through two parallel encoders, fuses their outputs by cross-attention, updates memory using a boundary-aware exemplar strategy, and performs final prediction with a decoupled Balanced Random Forest (BRF) classifier (Yang et al., 16 Aug 2025).

Component	Role	Main mechanism
Class-specific stream	Fine-grained discrimination	1D ResNet-18 + MOIA
Class-agnostic stream	Stable shared semantics	1D ResNet-18 + self-supervised continual learning
Fusion module	Guidance from coarse to fine	Multi-semantic cross-attention
Replay module	Forgetting mitigation	Boundary-Aware Exemplar Prioritization
Final classifier	Imbalance-robust prediction	Balanced Random Forest

Both streams use ResNet-18 backbones, but they are trained differently. The class-specific branch $C^t$0 is updated every session and optimized for discriminative learning. The class-agnostic branch $C^t$1 is trained with self-supervised continual learning so that it can function as a stable semantic anchor. The outputs are fused into a dual-semantic representation $C^t$2, which is used both for supervision and for guiding the class-specific branch (Yang et al., 16 Aug 2025).

This architecture is explicitly asymmetric. The class-agnostic branch is not merely an auxiliary regularizer; it is intended to preserve knowledge that the class-specific branch would otherwise overwrite. Conversely, the class-specific branch is the main carrier of new discriminative information. A plausible implication is that DGGN treats incremental diagnosis as a controlled interaction between stability and plasticity, with the guidance mechanism mediating between them (Yang et al., 16 Aug 2025).

3. Fine-grained and coarse-grained representation learning

The class-specific / fine-grained stream is a 1D ResNet-18 enhanced with the Multi-Order Interaction Aggregation (MOIA) module. MOIA takes an input feature

$C^t$3

splits channels into low-, mid-, and high-level groups, processes the latter two with dilated depthwise convolutions,

$C^t$4

and then concatenates and compresses them: $C^t$5 In the paper’s terminology, “multi-order” refers to contextual dependencies at different receptive-field scales rather than explicit polynomial interaction terms (Yang et al., 16 Aug 2025).

The class-specific branch is trained with three losses. First is supervised contrastive learning: $C^t$6 where $C^t$7. Second is feature-level knowledge distillation based on pairwise feature similarities, using

$C^t$8

The stated distillation loss is

$C^t$9

Third is a KL alignment term received from the fused representation, discussed below (Yang et al., 16 Aug 2025).

The class-agnostic / coarse-grained stream is also a 1D ResNet-18, but it is trained without class labels. At the base session it uses an InfoNCE objective: $t$0 with cosine similarity

$t$1

For incremental sessions, it adds cross-session alignment with a predictor $t$2: $t$3 The previous class-agnostic encoder $t$4 is frozen and used as a semantic anchor (Yang et al., 16 Aug 2025).

The stated rationale is that the class-agnostic stream remains stable because it is trained without class-discriminative supervision, is aligned to its previous version, and is used with StopGradient during fusion. The class-specific stream, by contrast, is intended to adapt rapidly to new classes. The paper’s CKA analysis is consistent with this interpretation: CA-CA inter-session similarities are reported as typically above 0.92, while CS-CS similarities are substantially lower (Yang et al., 16 Aug 2025).

4. Guidance mechanism, replay strategy, and decoupled classification

The core guidance module is multi-semantic cross-attention (MSCA). Given feature maps $t$5 and $t$6, DGGN first normalizes them: $t$7

$t$8

The class-specific feature provides the query,

$t$9

while both streams provide keys and values,

$D^t$0

$D^t$1

For each head, the fused representation is

$D^t$2

This makes the guidance explicitly uni-directional: the class-specific branch queries both itself and the class-agnostic branch, while the class-agnostic branch is protected from discriminative gradients by StopGradient (Yang et al., 16 Aug 2025).

The fused feature is supervised by

$D^t$3

and its predictive distribution $D^t$4 guides the class-specific branch distribution $D^t$5 through

$D^t$6

The class-specific branch loss is

$D^t$7

and the total objective is

$D^t$8

with implementation values

$D^t$9

(Yang et al., 16 Aug 2025).

To mitigate forgetting, DGGN uses Boundary-Aware Exemplar Prioritization (BAEP) rather than random or center-oriented replay. For a class candidate set $$D^1 = \left\{ (x_1, y_1), (x_2, y_2), \ldots, (x_i, y_i),\; y_i \in C^1 \right\},$$0 and class mean

$$D^1 = \left\{ (x_1, y_1), (x_2, y_2), \ldots, (x_i, y_i),\; y_i \in C^1 \right\},$$1

the selected exemplar is

$$D^1 = \left\{ (x_1, y_1), (x_2, y_2), \ldots, (x_i, y_i),\; y_i \in C^1 \right\},$$2

The memory budget is fixed at $$D^1 = \left\{ (x_1, y_1), (x_2, y_2), \ldots, (x_i, y_i),\; y_i \in C^1 \right\},$$3, with per-class quota

$$D^1 = \left\{ (x_1, y_1), (x_2, y_2), \ldots, (x_i, y_i),\; y_i \in C^1 \right\},$$4

after seeing $$D^1 = \left\{ (x_1, y_1), (x_2, y_2), \ldots, (x_i, y_i),\; y_i \in C^1 \right\},$$5 classes. This selection rule explicitly favors samples deviating from the class center, which the paper interprets as boundary-relevant and diversity-preserving (Yang et al., 16 Aug 2025).

Final prediction is decoupled from neural representation learning and performed by a Balanced Random Forest. For each tree $$D^1 = \left\{ (x_1, y_1), (x_2, y_2), \ldots, (x_i, y_i),\; y_i \in C^1 \right\},$$6,

$$D^1 = \left\{ (x_1, y_1), (x_2, y_2), \ldots, (x_i, y_i),\; y_i \in C^1 \right\},$$7

with balanced sampling between minority and majority classes. After training $$D^1 = \left\{ (x_1, y_1), (x_2, y_2), \ldots, (x_i, y_i),\; y_i \in C^1 \right\},$$8 trees, prediction is

$$D^1 = \left\{ (x_1, y_1), (x_2, y_2), \ldots, (x_i, y_i),\; y_i \in C^1 \right\},$$9

The paper’s stated reason for this decoupling is that few-shot incremental imbalance biases neural decision boundaries; BRF is intended to counter that bias more effectively than a standard softmax head (Yang et al., 16 Aug 2025).

5. Experimental evidence, ablations, and analytical findings

DGGN is evaluated on the TEP benchmark and a real-world MFF dataset. TEP provides 40 measured variables, 12 manipulated variables, and 21 fault types in total; the experiments use faults 1, 2, 4, 6, 7, 8, 12, 14, 18, plus normal class 0. MFF contains 24 process variables sampled at 1 Hz and six standard fault types; the experiments use faults 1, 2, 3, 4, plus normal class 0. The incremental setups are $t>1$0 classes for TEP and $t>1$1 for MFF (Yang et al., 16 Aug 2025).

The sample settings simulate few-shot imbalance. On TEP, normal training samples are 500, fault training samples are 48 per class in imbalanced mode and 20 per class in long-tailed mode, with 800 test samples for both normal and fault classes. On MFF, normal training samples are 200, fault training samples are 10 per class in imbalanced mode and 5 per class in long-tailed mode, again with 800 test samples for both normal and fault classes. Results are reported as classification accuracy and mean per-class accuracy on all seen classes, averaged over the last 10 checkpoints (Yang et al., 16 Aug 2025).

The main averages are uneven rather than uniformly dominant. On TEP imbalanced, DGGN reports 90.41 average accuracy versus 86.52 for SCLIFD. On TEP long-tailed, DGGN reports 85.13, while SCLIFD reports 86.01. On MFF imbalanced, DGGN reports 99.59 versus 97.94 for BiDistFSCIL and 97.86 for WaRP-CIFSL. On MFF long-tailed, DGGN reports 99.99 versus 98.87 for WaRP-CIFSL and 97.09 for BiDistFSCIL (Yang et al., 16 Aug 2025). This record is consistent with a neutral reading: DGGN is strongest on MFF and on TEP imbalanced, but not strictly best on every reported setting.

The ablations directly support the dual-granularity claim. Removing the class-agnostic branch is the largest degradation: average accuracy drops from 90.41 to 71.06 on TEP imbalanced, from 85.13 to 71.03 on TEP long-tailed, from 99.59 to 91.78 on MFF imbalanced, and from 99.99 to 87.40 on MFF long-tailed. Removing MSCA also causes major losses, including 90.41 \rightarrow 83.66 on TEP imbalanced and 85.13 \rightarrow 73.84 on TEP long-tailed. Removing KL-based knowledge transfer is milder on TEP imbalanced (90.41 \rightarrow 89.90) but substantial on TEP long-tailed (85.13 \rightarrow 76.93). Removing MOIA yields smaller but consistent declines, such as 90.41 \rightarrow 89.04 on TEP imbalanced and 99.99 \rightarrow 97.85 on MFF long-tailed (Yang et al., 16 Aug 2025).

The paper also reports classifier comparisons favoring BRF over FCC, SVM, and KNN. On TEP imbalanced, the averages are 80.64 for FCC, 88.10 for SVM, 90.26 for KNN, and 90.41 for BRF/DGGN. On TEP long-tailed, the corresponding values are 72.47, 78.42, 82.00, and 85.13 (Yang et al., 16 Aug 2025).

Analytically, the most direct evidence for the intended representation split is the CKA heatmap. Reported values include Agnostic-S1 vs Agnostic-S2: 0.98, Agnostic-S1 vs Agnostic-S5: 0.96, and generally CA-CA similarities above 0.92. By contrast, CS-CS similarities include Specific-S2 vs Specific-S3: 0.41 and Specific-S4 vs Specific-S5: 0.37, while cross-branch similarities such as Agnostic-S1 vs Specific-S1: 0.32 remain low (Yang et al., 16 Aug 2025). This supports the paper’s intended decomposition into a stable coarse-grained stream and a more session-sensitive fine-grained stream.

Implementation details are explicit. The model uses ResNet-18 in both streams, Adam, batch size 512, weight decay $t>1$2, temperature $t>1$3, learning rate 0.01, and 500 epochs. The memory buffer sizes are 100 and 40 on TEP for imbalanced and long-tailed settings, and 10 and 5 on MFF. The MOIA-based DGGN has about 9.91M parameters on TEP, compared with about 8.50M–8.54M for simpler DWConv variants, which suggests a modest parameter increase relative to the reported gains (Yang et al., 16 Aug 2025).

6. Broader use of “dual-granularity guidance” and the specific scope of DGGN

In the recent literature, several methods are DGGN-like in spirit but are not literally named “Dual-Granularity Guidance Network.” In embodied RL, DEG uses generated episodic guidance with coarse-grained exploration reward and fine-grained matching reward, but the paper explicitly states that it is “not literally a DGGN” and not a standalone guidance network (Liu et al., 13 Feb 2026). In agentic RL, D2Skill organizes reusable experience into task skills and step skills, a dual-granularity guidance memory rather than a monolithic network (Tu et al., 30 Mar 2026). In medical image classification, DiffMIC uses dual conditional guidance through global and local priors in a diffusion classifier (Yang et al., 2023). Similar two-level guidance ideas appear in infrared small target detection through DGSPNet (Wang et al., 24 Nov 2025), in graph learning through Cluster-GT and node/cluster interaction (Huang et al., 2024), in human trajectory prediction through MGTraj (Sun et al., 11 Sep 2025), and in controllable reasoning compression through CtrlCoT (Fan et al., 28 Jan 2026).

This makes the naming issue important. The explicit term DGGN in the supplied literature refers to the FSC-FD framework introduced in “Few-shot Class-incremental Fault Diagnosis by Preserving Class-Agnostic Knowledge with Dual-Granularity Representations” (Yang et al., 16 Aug 2025). Other works use dual-granularity guidance as a design principle, but not as the same method name. A common misconception is therefore to treat “DGGN” as a generic label for any coarse/fine two-stream model. The papers above suggest a narrower and more accurate usage: DGGN denotes the fault-diagnosis framework with class-specific/class-agnostic dual streams, MSCA fusion, BAEP replay, and BRF classification, whereas “dual-granularity guidance” is the broader methodological pattern spanning several domains (Yang et al., 16 Aug 2025).

From that perspective, DGGN occupies a specific place within the wider dual-granularity literature. Its distinctive feature is not merely that it has two streams, but that the coarse-grained stream is explicitly class-agnostic, self-supervised, session-stable, and protected by StopGradient, while the fine-grained stream is explicitly class-specific, MOIA-enhanced, and regularized by fused semantic guidance (Yang et al., 16 Aug 2025). This suggests that DGGN is best understood as a specialized continual-learning architecture in which coarse-grained knowledge functions as a preserved semantic anchor for few-shot class-incremental fault diagnosis, rather than as a generic name for all dual-granularity models.