Paper · 2021-12-16 · cs.LG
Intelligent Bearing Fault Diagnosis Method Combining Mixed Input and Hybrid CNN-MLP model
Rolling-bearing fault diagnosis from a sensor mounted on the shaft. Linear acceleration becomes 32x32 HHT images for a CNN; angular acceleration gives two spectral band powers (N1, N2) for an MLP. The two branches are concatenated to classify five conditions. Author-reported test accuracy is 99.6%, against 98% for the CNN alone and 81% for the MLP alone.
Convolutional neural networkMulti-layer perceptron (MLP)Multimodal fusionHilbert-Huang transformbearing-fault-diagnosismultimodal-inputsignal-processing
arXivPDFCodeMarkdown11 min read
AI-compiled by claude-haiku-5-5 on 2026-11-22, not yet reviewed by a human. Every claim cites its source; check it when a detail matters.
In plain words
Rolling bearings wear out and can bring a machine down, so spotting a fault early is worth a lot. This paper's sensor is a small wireless unit mounted on the spinning shaft itself, not on the machine's housing. It records two kinds of motion: straight-line acceleration and twisting (angular) acceleration.
The straight-line signal is turned into a small time-frequency picture, and a convolutional network reads that picture. The twisting signal is reduced to just two numbers, the signal energy near the shaft's two natural twisting frequencies, and a small network reads those two numbers. The two branches are joined and the model picks one of five conditions: normal, inner-race fault, outer-race fault, ball fault, or a combination of these. The authors report that the joined model beats either branch used alone.
TL;DR
- Input is one sensor on the shaft that gives linear and angular acceleration at 31.175 kHz §2.1 §3.2.
- Linear windows become 32x32 HHT images for a CNN; angular windows become two FFT band powers (N1, N2) for an MLP §2.1.
- The CNN and MLP outputs are concatenated and classified into five classes §3.3 tab. 2.
- Author-reported test accuracy: 99.6% (hybrid), 98% (CNN alone), 81% (MLP alone) on 4,185 test windows tab. 3 §4.
- The linked code is a dataset repository with notebooks. It does not contain the HHT or N1/N2 feature code, and its recorded outputs do not reproduce the paper's accuracy (see Paper vs code).
Problem
Rolling-bearing failures are a major share of failures in rotating machinery, so condition monitoring matters §1. Vibration-based diagnosis usually uses a sensor on the housing, and the usefulness of each signal depends on where that sensor sits §1. The authors argue that mounting on the shaft potentially increases sensitivity to defects, and that their data set is unique enough to publish §1 §4. The paper states its own scope limit: all experiments were run at a single shaft speed (1200 rpm, 20 Hz) §3.2 §4.
Method
The method has two stages: data organisation (conversion of the sensor signal to mixed inputs), then a hybrid model trained on those inputs §2 fig. 1.
Step 1: Mixed-input conversion §2.1 fig. 3.
- Split each signal into overlapping windows of 3,897 samples (125 ms, 2.5 shaft revolutions). Start a new window every 1,559 samples (50 ms, one revolution) §2.1.
- Linear acceleration: apply the Hilbert-Huang transform to each window, using only the first three intrinsic mode functions, and render a 32x32 time-frequency image §2.1.1.
- Angular acceleration: apply an FFT to the same window. N1 and N2 are the sums of the spectral components around the first and second torsional natural frequencies §2.1.2 fig. 7.
- Torsional frequencies come from a shock-response spectrum measured on the stationary shaft (hammer impact). The two peaks are at about 240 Hz and about 820 Hz §2.1.2 fig. 6.
- Normalise N1, N2 and image intensities to [0, 1] §2.2.
# Mixed-input conversion [§2.1] [fig. 3]
# Input: linear, angular: 1-D arrays sampled at 31.175 kHz [§3.2]
# Output: images (N, 32, 32, 3), numbers (N, 2)
def make_mixed_inputs(linear, angular):
# Overlapping windows: 3,897 samples, one every 1,559 samples [§2.1]
for linear_win, angular_win in sliding_windows(
linear, angular, length=3897, step=1559):
# Linear window -> HHT using the first three IMFs [§2.1.1]
image = hht_image(linear_win, num_imfs=3, size=32)
# Sum FFT power around each torsional peak; band edges not stated
# about 240 Hz peak [§2.1.2]
n1 = band_power(angular_win, around=PEAK_1_HZ)
# about 820 Hz peak [§2.1.2]
n2 = band_power(angular_win, around=PEAK_2_HZ)
yield image, [n1, n2]Step 2: Hybrid model §2.2 §3.3 tab. 2 fig. 8.
The CNN branch extracts features from the image, the MLP branch reads N1 and N2, and the two outputs are concatenated feature-wise before the classifier §2.2.
# Hybrid CNN-MLP with mixed input [§2.2] [tab. 2] [fig. 8]
# Input: image (B, 32, 32, 3); numbers (B, 2), both in [0, 1]
# Output: logits (B, 5), one per condition
def hybrid_forward(image, numbers):
# CNN branch: three conv+pool blocks, 3x3 kernels, zero padding [§3.3]
features = image
for filters in [16, 32, 64]:
features = relu(conv2d(features, filters, kernel=3, padding="same"))
features = max_pool(features, size=2)
# 4 x 4 x 64 map; the paper's table says the flatten layer has 256 units
features = flatten(features) # [tab. 2]
features = relu(dense(features, 16)) # D1 [tab. 2]
# Dropout 0.5 sits between D1 and D2 [§3.3]
features = dropout(features, rate=0.5)
features = relu(dense(features, 8)) # D2 [tab. 2]
# MLP branch on the two band powers [tab. 2]
numeric = relu(dense(numbers, 16))
numeric = relu(dense(numeric, 8))
# Concatenate both branches, then fuse with a dense layer [§3.3] [fig. 8]
fused = concatenate([features, numeric])
fused = relu(dense(fused, 8))
# Five output units; the output activation is not stated [§3.3]
return dense(fused, 5)Step 3: Training §3.3.
# Training loop [§3.3]
# Adam, learning rate 1e-4, batch size 20, categorical cross-entropy
def train(model, train_set, val_set):
# Initial epoch cap was 200 [§3.3]
best_val_loss = infinity
best_weights = None
epochs_since_best = 0
for epoch in range(200):
# One pass over the training set in batches of 20 [§3.3]
train_one_epoch(model, train_set, batch_size=20)
# The stopping rule uses validation loss [§3.3]
val_loss = evaluate_loss(model, val_set)
if val_loss < best_val_loss:
best_val_loss = val_loss
best_weights = copy_weights(model)
epochs_since_best = 0
else:
epochs_since_best += 1
# Patience of 50 epochs without improvement [§3.3]
if epochs_since_best == 50:
break
# Keep the weights with the lowest validation loss [§3.3]
return best_weightsData set §3.2 tab. 1. Five bearings were tested: one defect-free, three with a single fault (inner race, outer race, ball), and one with all three faults combined §3.1. Windows were split with scikit-learn train_test_split() §3.2.
| Fault type | Training | Validation | Test |
|---|---|---|---|
| Normal | 4036 | 705 | 839 |
| Inner race | 4056 | 696 | 828 |
| Outer race | 4018 | 746 | 816 |
| Ball | 3985 | 725 | 870 |
| Combined fault | 4062 | 686 | 832 |
| Total | 20157 | 3558 | 4185 |
The split ratio and random seed are not stated in the paper. The counts above sum to 27,900 HHT images §3.2 and give a test share of 15.0% (4185 / 27900, computed).
Hyperparameters
| Name | Value | Where | Notes |
|---|---|---|---|
| Sampling rate | 31.175 kHz | §2.1 §3.2 | |
| Shaft speed | 1200 rpm (20 Hz) | §3.2 | Single speed only |
| Window length | 3,897 samples (125 ms) | §2.1 | 2.5 revolutions |
| Window shift | 1,559 samples (50 ms) | §2.1 | One revolution |
| HHT image size | 32 x 32 | §2.1 §3.3 | Image has 3 channels §3.2 |
| HHT IMFs used | First 3 | §2.1.1 | "most effective" per the authors' experiments |
| Torsional peaks | about 240 Hz and about 820 Hz | §2.1.2 fig. 6 | From a shock-response spectrum |
| N1, N2 band edges | not stated | §2.1.2 fig. 7 | Figure shows the N2 band only |
| Input normalisation | [0, 1] | §2.2 | N1, N2 and pixel intensities |
| CNN conv blocks | 16, 32, 64 filters; 3x3; zero padding | tab. 2 §3.3 | 2x2 max pool after each |
| CNN dense | D1 16 units; dropout 0.5; D2 8 units | tab. 2 §3.3 | |
| Flatten | 256 units (as printed) | tab. 2 §3.3 | Conflicts with 4x4x64 = 1024; see Pitfalls |
| MLP branch | 16, then 8 units | tab. 2 | Input N1, N2 |
| Fusion | Concatenate, dense 8 | tab. 2 §3.3 | |
| Output | 5 units | tab. 2 §3.3 | Activation not stated |
| Activation | ReLU everywhere | §3.3 | |
| Optimiser | Adam, learning rate 1e-4 | §3.3 | Decay, betas and epsilon not stated |
| Batch size | 20 | §3.3 | |
| Loss | Categorical cross-entropy | §3.3 | |
| Epoch cap | 200 ("initially") | §3.3 | |
| Early-stopping patience | 50 epochs | §3.3 | Monitors validation loss |
| Checkpoint | Lowest validation loss | §3.3 | Hybrid: stopped at epoch 112, kept epoch 62 |
| Seed | not stated |
Results
Author-reported. Setup: one bearing rig at 1200 rpm, 4,185 test windows, five classes §3.2 tab. 1 tab. 3. Baselines are the same authors' CNN and MLP branches trained alone; the paper states the same stopping rule (50 epochs without a validation-loss decrease) for them §3.3 fig. 13. The paper does not say how many runs were averaged.
| Model | Accuracy | Source |
|---|---|---|
| Hybrid CNN-MLP | 99.6% (1.00 in table) | §3.3 §4 tab. 3 |
| CNN alone | 98% (0.98 in table) | tab. 3 |
| MLP alone | 81% (0.81 in table) | tab. 3 |
Hybrid confusion matrix: 15 errors in 4185 test windows §3.3 fig. 13. The CNN's main confusion is ball faults predicted as inner race (56 windows); the MLP confuses normal with ball and outer with combined §3.3 fig. 13.
Per-class precision, recall and F1 tab. 3:
| Class | Hybrid P | Hybrid R | Hybrid F1 | CNN P | CNN R | CNN F1 | MLP P | MLP R | MLP F1 |
|---|---|---|---|---|---|---|---|---|---|
| Normal | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 | 0.77 | 0.79 | 0.78 |
| Inner race | 0.99 | 1.00 | 0.99 | 0.93 | 0.96 | 0.94 | 0.95 | 0.94 | 0.95 |
| Outer race | 1.00 | 1.00 | 1.00 | 1.00 | 0.99 | 0.99 | 0.76 | 0.74 | 0.75 |
| Ball | 1.00 | 0.99 | 0.99 | 0.96 | 0.94 | 0.95 | 0.78 | 0.74 | 0.76 |
| Combined fault | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 | 0.78 | 0.83 | 0.80 |
Other reported figures: t-SNE of the HHT images (after PCA to 506 components) shows three visual clusters, with inner-race and ball faults overlapping §3.2 fig. 11 fig. 12.
Paper vs code
The only linked repository is github.com/susu-cm/bearings-dataset, which the paper cites as its dataset §4 [ref. 35]. Its owner is a user account, not an organisation. The README says the data was "assembled and prepared on behalf of South Ural State University" and cites a different ScienceDirect article for the dataset (pii S0888327022005714). The repository has no LICENSE file. It holds .mat files at 18 Hz and 20 Hz and one folder LPC-NN/ with three notebooks. The hybrid model is in Hybrid CNN-MLP_LPC-NN_TDSIs-NN_SUSU-Example 4_3.ipynb; the other two notebooks are named for CWRU and were not checked. Commit c31985c9c1c0058f97dcb94a648e962fa665de89.
- Feature extraction is missing. No HHT, EMD, FFT or N1/N2 code appears in the hybrid notebook. The two CWRU notebooks were not checked for it. The notebook loads pre-made PNG images and
TxtInfo_*.txtfiles from local Windows paths Hybrid CNN-MLP_LPC-NN_TDSIs-NN_SUSU-Example 4_3.ipynb Hybrid CNN-MLP_LPC-NN_TDSIs-NN_SUSU-Example 4_3.ipynb. Those files are not in the repository, so the window, shift, IMF and band-edge choices cannot be checked against code. - Split sizes differ. The notebook splits the training file with
test_size=0.1, random_state=42Hybrid CNN-MLP_LPC-NN_TDSIs-NN_SUSU-Example 4_3.ipynb. Its printed shapes are 1939 train, 216 validation and 920 test windows Hybrid CNN-MLP_LPC-NN_TDSIs-NN_SUSU-Example 4_3.ipynb. Paper Table 1 reports 20157 / 3558 / 4185. - Scaler is refitted per split. The notebook calls
fit_transformon the training set, thenfit_transformagain on the validation set, thentransformon the test set using the validation-fitted scaler Hybrid CNN-MLP_LPC-NN_TDSIs-NN_SUSU-Example 4_3.ipynb. The paper says only that inputs are "normalized to the range of [0,1]" §2.2. - Early stopping differs. The paper uses patience 50, a 200-epoch initial cap, and validation loss §3.3. The code's ten-run loop uses patience 100, 500 epochs, and monitors
val_accuracy, with the checkpoint also keyed toval_accuracyHybrid CNN-MLP_LPC-NN_TDSIs-NN_SUSU-Example 4_3.ipynb Hybrid CNN-MLP_LPC-NN_TDSIs-NN_SUSU-Example 4_3.ipynb. A separate single-run cell uses patience 10 Hybrid CNN-MLP_LPC-NN_TDSIs-NN_SUSU-Example 4_3.ipynb. - Optimiser decay is in code only. Adam uses
learning_rate=1e-4, decay=1e-3/200Hybrid CNN-MLP_LPC-NN_TDSIs-NN_SUSU-Example 4_3.ipynb. The paper gives only the learning rate. - Batch normalisation is in code only. Each CNN conv block has ReLU, batch normalisation and max pooling; batch normalisation is also after the first dense layer, and the CNN uses
padding="same"Hybrid CNN-MLP_LPC-NN_TDSIs-NN_SUSU-Example 4_3.ipynb. The paper describes none of these, though it does say zero padding is used §3.3. - Flatten size differs from the table. The code's flatten layer has 1024 units Hybrid CNN-MLP_LPC-NN_TDSIs-NN_SUSU-Example 4_3.ipynb, matching 4x4x64, not the 256 printed in Table 2.
- Output head. The code's final layer is 5 linear units trained with
CategoricalCrossentropy(from_logits=True), so there is no softmax layer Hybrid CNN-MLP_LPC-NN_TDSIs-NN_SUSU-Example 4_3.ipynb Hybrid CNN-MLP_LPC-NN_TDSIs-NN_SUSU-Example 4_3.ipynb. The paper leaves the output activation unstated. - Results do not reproduce. The notebook's ten-run mean test accuracy is 0.9576, with per-run values from 0.945 to 0.970 Hybrid CNN-MLP_LPC-NN_TDSIs-NN_SUSU-Example 4_3.ipynb. The paper's 99.6% comes from a 4,185-window test set that the notebook does not use. The notebook's results cell evaluates one "last run" Hybrid CNN-MLP_LPC-NN_TDSIs-NN_SUSU-Example 4_3.ipynb.
- Other notebook cells are other methods. The notebook also contains LPC-NN and TDSIs-NN cells, which the notebook header ties to the paper's other examples. Their architectures are not described in this paper. The standalone CNN and MLP baselines in Table 3 are not located in the repository.
- Data files in the repository. The
.matfiles are the raw acceleration records for both 18 Hz and 20 Hz. The paper uses only the 20 Hz (1200 rpm) data §3.2.
Pitfalls
- Band edges for N1 and N2 are not stated. The text says "the sum of the FFT components around the first and second frequency peaks" §2.1.2, and Figure 7 shows one shaded band for N2. Without edges, reproducing N1 and N2 is a guess. Choose edges from the shock-response spectrum and report them.
- Peak frequencies and shaft size are inconsistent. The paper gives the torsional peaks as "approximately 240 Hz and 820 Hz" §2.1.2. The shock-response spectrum is measured on a 3/4-inch shaft §2.1.2 fig. 6, but the experimental rig uses a 1-inch shaft §3.1. Which shaft the N1 and N2 bands came from is not clear.
- HHT details are missing. The paper does not give the EMD sifting stop criterion, the envelope interpolation, the frequency and time bin counts, the colour map, or how the 32x32 grid maps to frequency and time. Each choice changes the image.
- Image channels. The paper says each image has 32x32x3 = 3072 values §3.2, so the image is treated as three-channel. How the three channels are formed is not stated.
- Flatten size. Table 2 prints 256 units for the flatten layer, but the text says 64 maps of size 4x4 §3.3, which is 1024 values. Use 1024 unless the table is confirmed.
- Validation-loss stopping vs accuracy stopping. The paper says stopping and checkpointing use validation loss §3.3. The code uses validation accuracy Hybrid CNN-MLP_LPC-NN_TDSIs-NN_SUSU-Example 4_3.ipynb. The two can pick different epochs.
- Single speed and single rig. All experiments use one shaft speed and one rig §3.2 §4. The authors list other speeds as future work §4. Expect the numbers to change with speed.
- Each error is worth about 0.024 points. The hybrid model makes 15 errors in 4185 windows §3.3, so one more or one fewer error moves the reported 99.6% by about 0.024 percentage points. Compare models on counts as well as percentages.
- Number of runs and variance. The paper does not say how many runs were averaged or give seeds or spread §3.3. The notebook's ten runs vary from 0.945 to 0.970 Hybrid CNN-MLP_LPC-NN_TDSIs-NN_SUSU-Example 4_3.ipynb. Report mean and spread.
- Overlapping windows across the split. Consecutive windows share 2,338 samples (3,897 minus 1,559) §2.1. The paper splits the window set with
train_test_split()§3.2, which is a per-window split, so near-duplicate windows from the same recording can fall on both sides. The paper does not say whether the split respects recordings. - Dataset README points to another article. The repository README cites a ScienceDirect article for the dataset (pii S0888327022005714), which is not this arXiv v1. Use this paper's own text for the dataset description.
- Licence of the code. The repository has no LICENSE file, so no code can be quoted verbatim from it. The citations above are for comparison only.
Lineage
- Builds on: Hilbert-Huang transform [ref. 36], and the wireless acceleration sensor described in ref. [32], which the paper cites for the sensor's construction §2.1. The paper calls HHT-CNN "one of the state-of-the-art bearing fault diagnosis schemes" §1 and cites ref. [27] for HHT use in bearing analysis §2.1.1.
- Compared against: the same paper's CNN branch and MLP branch trained alone §3.3 tab. 3.
- Superseded by: not stated.
Source and license
Adapted from “Intelligent Bearing Fault Diagnosis Method Combining Mixed Input and Hybrid CNN-MLP model” by V. Sinitsin, O. Ibryaeva, V. Sakovskaya, V. Eremeeva (arXiv:2112.08673), used under CC BY 4.0. Changes: condensed and restructured for implementation, with tables, equations and algorithms taken from the paper and its code, each cited to its source. Not affiliated with or endorsed by the authors. Code citations refer to https://github.com/susu-cm/bearings-dataset at commit c31985c9c1c0 (none (no LICENSE file in repository)). Any quoted code is shown under that license.