In this work, I address a practical and difficult problem in industrial robot condition monitoring: joint motor bearing faults that are hidden by strong background noise and variable operating conditions. In an industrial robot, the joint motor bearing is a critical component, and its degradation can directly reduce positioning accuracy, increase vibration, and eventually cause unplanned downtime. Traditional blind deconvolution methods can extract periodic fault impulses, but they often require manually selected bandpass parameters and become unstable when the fault characteristic frequency drifts with speed and load. Deep learning methods can classify faults from raw signals, but their black-box nature limits physical interpretability and their performance degrades sharply when the signal-to-noise ratio is low. To overcome these limitations, I propose a time-frequency optimized blind deconvolution framework, denoted TFO-BD, for fault diagnosis and prognosis of an industrial robot joint motor bearing. The proposed method explicitly embeds a time-domain quadratic convolution filter and a frequency-domain linear filter into a deep learning architecture. The time-domain filter suppresses non-periodic noise, while the frequency-domain filter enhances the fault characteristic frequency information. A physical-informed loss function that combines kurtosis constraints, frequency-domain energy constraints, and a cross-entropy term enables end-to-end joint optimization of the blind deconvolution parameters and the classification network. On this basis, I build a multi-step time-series prediction model using a gated recurrent unit network to forecast fault evolution and provide early warning feedback. The diagnosis and prediction modules form a closed-loop system for an industrial robot.

1. Problem Formulation and Motivation
An industrial robot joint system operates under high load, repeated motion, and changing speed. These conditions make the joint motor bearing susceptible to wear, fatigue, and plastic deformation. When a local defect appears on the inner ring or outer ring, the bearing generates periodic impulses. The repetition rate is related to the fault characteristic frequency, which depends on bearing geometry and rotational speed. However, in a real industrial robot, the measured current signal is a mixture of fault impulses, structural resonances, electromagnetic interference, and broadband noise. The periodic fault components are often weak and can be completely submerged at low signal-to-noise ratios. Therefore, the central task is to recover a fault-related source signal from the measured current signal and then classify the fault type.
I formulate the signal transmission process as a convolution model. Let the measured signal be \(x \in \mathbb{R}^{N}\), the fault source signal be \(d \in \mathbb{R}^{N}\), and the noise be \(n_{\mathrm{noise}} \in \mathbb{R}^{N}\). The measured signal can be written as
$$
x = d * h_d + n_{\mathrm{noise}} * h_n
$$
where \(h_d\) and \(h_n\) are transfer functions, and \(*\) denotes convolution. The purpose of blind deconvolution is to find a filter \(f \in \mathbb{R}^{L}\) such that the filtered signal \(y \in \mathbb{R}^{N}\) approximates the fault source:
$$
y = x * f = (d * h_d + n_{\mathrm{noise}} * h_n) * f \approx d
$$
The above expression is an approximate linear description of the actual industrial robot joint system. In practice, the propagation path is not strictly linear, and the noise term includes periodic harmonics caused by mechanical rotation, random impulses from external impacts, and broadband background noise. This mismatch motivates a time-frequency joint design rather than a purely time-domain or purely frequency-domain blind deconvolution filter.
For a rolling bearing, the fault characteristic frequency can be expressed from geometry and rotational speed. For an outer-ring defect, the ball-pass frequency can be written as
$$
f_{\mathrm{BPFO}} = \frac{N_b}{2} f_r \left(1 – \frac{d_b}{d_p}\cos\theta\right)
$$
and for an inner-ring defect, it can be written as
$$
f_{\mathrm{BPFI}} = \frac{N_b}{2} f_r \left(1 + \frac{d_b}{d_p}\cos\theta\right)
$$
where \(N_b\) is the number of balls, \(f_r\) is the rotational frequency, \(d_b\) is the ball diameter, \(d_p\) is the pitch diameter, and \(\theta\) is the contact angle. In an industrial robot, \(f_r\) changes with joint motion, so the fault characteristic frequency is non-stationary. This is the main reason why a fixed bandpass filter or a hand-tuned blind deconvolution filter cannot remain reliable across operating conditions.
To clarify the notation used throughout this work, I summarize the main symbols in Table 1.
| Symbol | Description |
|---|---|
| \(x\) | Measured current signal from the industrial robot joint motor |
| \(d\) | Fault source signal related to periodic impulses |
| \(h_d, h_n\) | Transfer functions of the fault path and noise path |
| \(n_{\mathrm{noise}}\) | Noise and interference components |
| \(f\) | Blind deconvolution filter |
| \(y\) | Filtered signal after blind deconvolution |
| \(\mathrm{FCF}\) | Fault characteristic frequency |
| \(L_t\) | Time-domain kurtosis loss |
| \(L_f\) | Frequency-domain energy loss |
| \(L_c\) | Cross-entropy classification loss |
| \(\sigma_c, \sigma_t, \sigma_f\) | Learnable uncertainty weights for the three loss terms |
2. Data Description and Preprocessing
I evaluate the proposed method on current signals collected from servo motor bearings of an industrial robot. The dataset contains multiple health states and two representative operating conditions. The current signal is selected because it can be acquired non-invasively from the motor drive and is convenient for online monitoring in an industrial robot. Each record contains 250,000 current samples within a 4 s acquisition window. The data are repeated 20 times under the same condition to improve statistical reliability. The operating conditions are summarized in Table 2.
| Operating case | Speed / rpm | Torque / N·m | Load / N | Record label |
|---|---|---|---|---|
| Case I | 900 | 0.7 | 1,000 | N09_M07_F10 |
| Case II | 1,500 | 0.1 | 1,000 | N15_M01_F10 |
The dataset includes normal data and 13 fault classes. The fault labels use a multidimensional composite coding rule. \(\mathrm{OR}\) and \(\mathrm{IR}\) denote outer-ring and inner-ring faults, respectively. \(\mathrm{S}\), \(\mathrm{R}\), and \(\mathrm{M}\) denote single, repeated, and multiple damage. \(\mathrm{D1}\) to \(\mathrm{D3}\) denote damage levels. \(\mathrm{T}\) and \(\mathrm{P}\) denote fatigue and plastic deformation. Each fault class is identified by a four-element combination that describes spatial distribution, morphology, severity, and physical cause. This coding system is useful because it allows the diagnosis network to distinguish not only the fault location but also the damage pattern and severity.
For preprocessing, I first normalize each feature to the interval \([0,1]\). The normalization formula is
$$
x’_i = \frac{x_i – x_{\min}}{x_{\max} – x_{\min}}
$$
where \(x_{\min}\) and \(x_{\max}\) are the minimum and maximum values of the same feature in a given state. I then segment the current signal into fixed-length samples. The segment length is 1,024, and the stride is also 1,024. This produces a dataset of size \(15,200 \times 1,024\). The segmentation step ensures that each sample has a consistent length for one-dimensional convolutional neural networks and allows graphics processing unit acceleration.
Before classification, I extract a set of physical features from each current segment. These features include the root mean square, variance, kurtosis, waveform index, spectral energy, wavelet packet decomposition energy at levels 0 and 1, and spectral peak. The root mean square reflects the overall energy level. The variance describes the fluctuation intensity. Kurtosis is sensitive to impulsive components, which are closely related to bearing defects. The waveform index describes the shape of the time-domain waveform. Spectral energy and spectral peak capture the frequency-domain distribution. Wavelet packet energy at different levels provides a time-frequency representation of transient components. The candidate features are listed in Table 3.
| Feature | Physical meaning | Mathematical expression |
|---|---|---|
| Root mean square | Overall energy of the current signal | \(\mathrm{RMS} = \sqrt{\frac{1}{N}\sum_{i=1}^{N} x_i^2}\) |
| Variance | Dispersion of the current signal | \(\sigma^2 = \frac{1}{N}\sum_{i=1}^{N}(x_i – \bar{x})^2\) |
| Kurtosis | Sensitivity to impulsive fault components | \(K = \frac{\sum_{i=1}^{N}(x_i – \bar{x})^4}{N\sigma^4}\) |
| Waveform index | Shape factor of the time-domain waveform | \(W = \frac{\mathrm{RMS}}{\frac{1}{N}\sum_{i=1}^{N}|x_i|}\) |
| Spectral energy | Energy distribution in the frequency domain | \(E = \sum_{k=1}^{K}|X(k)|^2\) |
| Wavelet packet energy | Energy of transient components at different scales | \(E_{j} = \sum_{m} |c_{j,m}|^2\) |
| Spectral peak | Dominant frequency amplitude | \(P = \max_{k}|X(k)|\) |
To avoid redundant or weakly discriminative features, I perform feature selection based on class separation distance. Let \(N_c\) be the number of health states, and let \(n\) be the number of signal samples for each state. For feature \(j\), the mean value of class \(c\) is
$$
m_{jc} = \frac{1}{n}\sum_{i=1}^{n} x_{ijc}
$$
The separation distance of feature \(j\) is calculated as
$$
d_j = \frac{1}{n N_c}\sum_{c=1}^{N_c}\sum_{i=1}^{n} (x_{ijc} – m_{jc})^2
$$
Then I normalize the separation distance by the maximum distance:
$$
\bar{d}_j = \frac{d_j}{\max(d)}
$$
I set the threshold to 0.7 and retain features whose normalized separation distance exceeds this threshold. The selected features form the input to the subsequent diagnosis and prediction models. This step reduces the influence of irrelevant noise and improves the stability of the industrial robot fault diagnosis pipeline.
3. Time-Frequency Optimized Blind Deconvolution
The proposed TFO-BD method is designed to solve two problems simultaneously: suppressing non-periodic noise in the time domain and enhancing periodic fault characteristic frequency components in the frequency domain. Unlike conventional blind deconvolution, which extracts features first and then trains a classifier separately, TFO-BD embeds the blind deconvolution filters directly into the deep learning graph. Therefore, the filters and the classifier share parameters and are optimized by a unified objective.
3.1 Time-Domain Quadratic Convolution Filter
High-order neural networks, also called polynomial neural networks, can model nonlinear interactions. A polynomial network can be written as
$$
y_i = \sigma\left(\prod_{j}\left(\sum_{k} w_{ijk} x_k + b_{ij}\right)\right)
$$
where \(w_{ijk}\) and \(b_{ij}\) are the weight and bias of the \(j\)-th polynomial product term for the \(i\)-th output node, \(\sigma(\cdot)\) is the activation function, and \(x_k\) is the \(k\)-th input feature. Although higher-order polynomial functions increase representation ability, they also increase computational complexity and make training unstable. For this reason, I restrict the polynomial order to two and use a quadratic neural network. A standard linear neuron is
$$
y = \sigma(W^T x + b)
$$
Replacing the linear function with a quadratic function gives
$$
y = \sigma\left((W_1^T x + b_1)(W_2^T x + b_2) + W_3^T (x \odot x) + b_3\right)
$$
where \(W_1\), \(W_2\), and \(W_3\) are learnable weight matrices, \(b_1\), \(b_2\), and \(b_3\) are bias vectors, and \(\odot\) denotes element-wise multiplication. This quadratic neuron can capture nonlinear modulation effects caused by bearing impacts and structural resonances. In the context of an industrial robot, these nonlinear effects are common because the joint transmission path changes with posture and load.
To stabilize training, I use a re-linearization initialization strategy. The first-layer weights and biases are initialized as
$$
W_1 \sim N(0, \sigma^2), \quad \sigma = \frac{1}{\sqrt{32 k_n}}
$$
$$
b_1 \sim U(-B, B), \quad B = \frac{1}{k_n}
$$
and the high-order terms are initially set to zero:
$$
W_2 = 0, \quad W_3 = 0, \quad b_2 = 1, \quad b_3 = 0
$$
where \(N(0,\sigma^2)\) is a Gaussian distribution, \(U(-B,B)\) is a uniform distribution, and \(k_n\) is the kernel size of \(W_1\). This initialization makes the quadratic network behave like a first-order linear system at the beginning of training. As training proceeds, the high-order weights grow slowly, and the network gradually learns nonlinear feature extraction. In my experiments, this “first linear, then nonlinear” behavior significantly improves convergence stability under strong noise.
The time-domain filter consists of two symmetric quadratic convolutional neural network layers. The first layer maps the single-channel input to 16 feature channels. This operation can be interpreted as a basis expansion of different fault frequency bands. The second layer fuses the 16 channels back to one channel through an inverse operation, ensuring that the output dimension is the same as the input dimension. I choose 16 channels because it provides a good balance between feature completeness and model complexity. A larger number of channels would increase the risk of overfitting and inference delay, while a smaller number would not cover the main fault frequency bands of an industrial robot joint bearing.
For the time-domain blind deconvolution filter, I use kurtosis as the objective. The time-domain loss is defined as
$$
L_t = – \frac{\sum_{n=1}^{N} y(n)^4}{\left(\sum_{n=1}^{N} y(n)^2\right)^2}
$$
where \(y(n)\) is the output of the time-domain filter. Minimizing this loss maximizes the kurtosis of the filtered signal, which promotes impulsive and periodic fault features.
3.2 Frequency-Domain Linear Filter
The frequency-domain filter is implemented as a fully connected layer. Its purpose is to enhance the fault characteristic frequency and suppress frequency components that are not related to bearing defects. Because the fault characteristic frequency of an industrial robot joint bearing drifts with rotational speed, the frequency-domain filter should be learned jointly with the downstream classifier rather than fixed in advance. I extract the output of the fully connected layer and compute a subspace energy spectrum. The frequency-domain loss encourages energy concentration at the fault characteristic frequency:
$$
L_f = -\frac{\sum_{k \in \mathcal{F}_{\mathrm{FCF}}} |Y(k)|^2}{\sum_{k=1}^{K} |Y(k)|^2}
$$
where \(\mathcal{F}_{\mathrm{FCF}}\) denotes the set of frequency bins associated with the fault characteristic frequency, and \(Y(k)\) is the Fourier transform of the filtered signal. In addition, I include an L2/L4 sparsity-promoting term to improve the concentration of periodic components:
$$
L_f = -\alpha \frac{\sum_{k \in \mathcal{F}_{\mathrm{FCF}}} |Y(k)|^2}{\sum_{k=1}^{K} |Y(k)|^2} + \beta \frac{\|Y\|_2^2}{\|Y\|_4^2 + \epsilon}
$$
where \(\alpha\) and \(\beta\) are balancing coefficients, and \(\epsilon\) is a small constant for numerical stability. The first term enhances the fault characteristic frequency, while the second term encourages a sparse frequency representation. This design is especially useful for an industrial robot because the measured current signal often contains multiple harmonic components from the motor drive and the control loop.
3.3 Classification Loss and Joint Optimization
After the time-domain and frequency-domain filters, the filtered signal is passed to a one-dimensional deep learning classifier. The classifier can be a deep residual network, a convolutional neural network, a Transformer, or a lightweight network such as MobileNetV3. I use a deep convolutional neural network with a wide first-layer kernel as a compact and effective classifier. The classification loss is the cross-entropy loss:
$$
L_c = – \sum_{i} p_i \log q_i
$$
where \(p_i\) is the true label probability of sample \(i\), and \(q_i\) is the predicted probability. To jointly optimize the blind deconvolution filters and the classifier, I define the overall loss as
$$
L = L_c + L_f + L_t
$$
Because the three loss terms have different scales and convergence speeds, I adopt an uncertainty-based weighting strategy. Each loss term is associated with a learnable standard deviation \(\sigma_c\), \(\sigma_t\), and \(\sigma_f\). The weighted loss is
$$
L = \frac{1}{\sigma_c^2}L_c + \frac{1}{\sigma_t^2}L_t + \frac{1}{\sigma_f^2}L_f + \log\sigma_c + \log\sigma_t + \log\sigma_f
$$
Each \(\sigma\) is initialized to 0.5 and optimized during training. If a \(\sigma\) becomes too large, the last logarithmic term penalizes it. This strategy does not guarantee a globally optimal balance, but it allows each loss term to decrease smoothly and prevents any single term from dominating the optimization. The loss components are summarized in Table 4.
| Loss component | Domain | Purpose | Expression |
|---|---|---|---|
| \(L_t\) | Time domain | Enhance impulsive and periodic fault features | \(-\frac{\sum y(n)^4}{(\sum y(n)^2)^2}\) |
| \(L_f\) | Frequency domain | Enhance fault characteristic frequency energy | \(-\frac{\sum_{k\in\mathcal{F}_{\mathrm{FCF}}}|Y(k)|^2}{\sum_k |Y(k)|^2}\) |
| \(L_c\) | Label space | Minimize classification error | \(-\sum_i p_i \log q_i\) |
| \(L\) | Joint | End-to-end optimization of filters and classifier | \(\frac{1}{\sigma_c^2}L_c + \frac{1}{\sigma_t^2}L_t + \frac{1}{\sigma_f^2}L_f + \log\sigma_c + \log\sigma_t + \log\sigma_f\) |
The proposed TFO-BD method differs from existing neural blind deconvolution and physics-informed diagnosis methods in three important ways. First, it jointly models time-domain nonlinear impulse amplification and frequency-domain periodic feature enhancement. Second, it embeds fault labels directly into the blind deconvolution learning process through the joint loss. Third, its parameters have clear physical meanings: the time-domain quadratic term corresponds to nonlinear modulation of fault impacts, and the frequency-domain linear filter corresponds to the frequency response of the structural transmission path. These properties make TFO-BD suitable for an industrial robot operating in a strong-noise environment.
4. Diagnosis and Prognosis Framework
The complete framework contains a diagnosis branch and a prognosis branch. The diagnosis branch uses TFO-BD as a plug-and-play denoising and feature enhancement module before a one-dimensional classifier. The prognosis branch uses a recurrent neural network to predict future current signals. The predicted current signal is then fed back to the diagnosis branch, forming a prediction-diagnosis closed loop for an industrial robot.
4.1 Diagnosis Branch
In the diagnosis branch, the current signal first passes through the time-domain quadratic convolution filter and the frequency-domain linear filter. The filters ensure that the output has the same dimension as the input. Therefore, any one-dimensional classifier can be connected after TFO-BD without changing its input shape. I evaluate four mainstream backbone networks: ResNet18, WDCNN, Transformer, and MobileNetV3. This design demonstrates the generality of TFO-BD for industrial robot fault diagnosis.
4.2 Prognosis Branch with Gated Recurrent Unit
Fault diagnosis identifies the current health state, while prognosis predicts the future trend and provides early warning. I use a gated recurrent unit network because it can model temporal dependencies with fewer parameters than a long short-term memory network. The gated recurrent unit is governed by the following equations:
$$
z_t = \sigma(W_z [h_{t-1}, x_t] + b_z)
$$
$$
r_t = \sigma(W_r [h_{t-1}, x_t] + b_r)
$$
$$
\tilde{h}_t = \tanh(W_h [r_t \odot h_{t-1}, x_t] + b_h)
$$
$$
h_t = (1 – z_t) \odot h_{t-1} + z_t \odot \tilde{h}_t
$$
where \(z_t\) is the update gate, \(r_t\) is the reset gate, \(\tilde{h}_t\) is the candidate hidden state, and \(h_t\) is the final hidden state. The network contains a recurrent layer followed by dropout, another recurrent layer, and a fully connected layer. The dropout layer reduces overfitting when the number of model parameters is large but the training samples are limited.
To evaluate prediction accuracy, I use the mean squared error, root mean squared error, and mean absolute error:
$$
L_{\mathrm{MSE}} = \frac{1}{N}\sum_{i=1}^{N}(y_i – \hat{y}_i)^2
$$
$$
L_{\mathrm{RMSE}} = \sqrt{L_{\mathrm{MSE}}}
$$
$$
L_{\mathrm{MAE}} = \frac{1}{N}\sum_{i=1}^{N}|y_i – \hat{y}_i|
$$
where \(y_i\) is the measured current value at time step \(i\), \(\hat{y}_i\) is the predicted current value, and \(N\) is the total number of test samples. For periodic signals, single-step prediction has large variance. Therefore, I use a cyclic multi-step prediction strategy. A fixed step length is selected, and the prediction is performed iteratively. The last predicted value in each cycle is recorded as the reference for the next cycle. Through comparison experiments over a step range from 60 to 200, I find that a step length of 120 gives the best balance between accuracy and stability. When the step length is too small, the prediction is sensitive to high-frequency noise and the computation is inefficient. When the step length is too large, the single-step error accumulates and the predicted trend may diverge. The step length selection is summarized in Table 5.
| Step length range | Observed behavior | Recommendation |
|---|---|---|
| Less than 60 | Sensitive to high-frequency noise; low computational efficiency | Not recommended |
| 60 to 200 | Stable for most cases; accuracy depends on signal periodicity | Use 120 as the preferred value |
| Greater than 200 | Single-step error accumulates; trend may diverge | Not recommended |
For periodic prediction, I extract signal segments at multiple cycle nodes. One record contains 250,000 samples. Visual analysis shows that one period contains about 1,000 samples, so each record can provide 250 periodic signal groups. The training-to-testing ratio is set to 7:3. This setup allows the model to learn the periodic pattern and evaluate its generalization ability.
4.3 Closed-Loop Prediction and Diagnosis
The prognosis branch predicts the next stage of current signal. The predicted signal is then passed to the diagnosis branch, which outputs the possible fault mode. In this way, the system can provide early warning before the fault becomes severe. For an industrial robot, this closed-loop mechanism is valuable because it connects condition prediction with fault classification and enables maintenance decisions to be made in advance.
5. Experimental Setup and Results
To simulate a strong-noise industrial robot environment, I add Gaussian white noise to the current signals. The noise level is controlled by the signal-to-noise ratio. I evaluate the proposed method at \(-4\) dB, which is a challenging condition for bearing fault diagnosis. The experimental configuration is listed in Table 6.
| Parameter | Setting |
|---|---|
| Noise type | Gaussian white noise |
| Signal-to-noise ratio | \(-4\) dB |
| Maximum training epochs | 200 |
| Batch size | 128 |
| Learning rate candidates | 0.1, 0.3, 0.5, 0.8 |
| Optimizer | Stochastic gradient descent with dynamic learning rate |
| Number of independent runs | 10 |
| Evaluation metrics | False positive rate and F1 score |
The classification results are shown in Table 7. The proposed TFO-BD consistently improves the performance of all backbone networks. For the N15M07F04 case, MobileNetV3 improves from 86.95% to 94.50%, ResNet18 improves from 40.57% to 77.28%, WDCNN improves from 81.59% to 86.52%, and Transformer improves from 83.68% to 88.37%. For the N09M07F10 case, MobileNetV3 improves from 71.86% to 74.49%, ResNet18 improves from 63.39% to 74.99%, WDCNN improves from 50.75% to 69.06%, and Transformer improves from 61.15% to 72.34%. On average, when the signal-to-noise ratio is \(-4\) dB, the F1 score of TFO-BD reaches 76.60%, while the original networks achieve only 63.07%.
| Operating case | Backbone network | Original F1 / % | Improved F1 / % | Gain / percentage points |
|---|---|---|---|---|
| N15M07F04 | MobileNetV3 | 86.95 | 94.50 | 7.55 |
| N15M07F04 | ResNet18 | 40.57 | 77.28 | 36.71 |
| N15M07F04 | WDCNN | 81.59 | 86.52 | 4.93 |
| N15M07F04 | Transformer | 83.68 | 88.37 | 4.69 |
| N09M07F10 | MobileNetV3 | 71.86 | 74.49 | 2.63 |
| N09M07F10 | ResNet18 | 63.39 | 74.99 | 11.60 |
| N09M07F10 | WDCNN | 50.75 | 69.06 | 18.31 |
| N09M07F10 | Transformer | 61.15 | 72.34 | 11.19 |
The results demonstrate that TFO-BD is especially effective when the original backbone is vulnerable to noise. For example, ResNet18 improves by 36.71 percentage points in the N15M07F04 case, and WDCNN improves by 18.31 percentage points in the N09M07F10 case. The reason is that the time-domain quadratic filter suppresses non-periodic noise, while the frequency-domain linear filter strengthens the fault characteristic frequency. The physical-informed loss also guides the filter to preserve fault-related components rather than random noise. These effects are critical for an industrial robot because the joint motor current signal is often contaminated by electromagnetic interference and time-varying load disturbances.
In the \(-4\) dB noise condition, the training and validation loss curves of the TFO-BD plus WDCNN model show some separation. The training loss continues to decrease, while the validation loss fluctuates. This behavior indicates a potential overfitting risk. However, the time-frequency filter and the physical constraints still allow the model to maintain core fault classification ability. The WDCNN F1 score increases from 50.75% to 69.06%. In future work, noise adaptation mechanisms such as dynamic frequency band selection or adversarial training can be introduced to reduce the training-validation gap and improve convergence stability.
6. Prediction Results and Model Comparison
I compare three recurrent neural network structures for fault prediction: a standard recurrent neural network, a long short-term memory network, and a gated recurrent unit network. The comparison is performed on the same current signal dataset. The gated recurrent unit network achieves the best overall performance. Its predicted curve is closest to the true current signal, and its error region is the smallest. The standard recurrent neural network has the largest error, while the long short-term memory network performs between the two. The qualitative comparison is summarized in Table 8.
| Prediction model | Accuracy | Stability | Training efficiency | Overall recommendation |
|---|---|---|---|---|
| Standard RNN | Lowest | Low | Moderate | Not preferred |
| LSTM | Moderate | Moderate | Lower than GRU | Useful but heavier |
| GRU | Highest | Highest | High | Preferred for industrial robot prognosis |
The mean squared error distribution further confirms this conclusion. The gated recurrent unit model has the lowest median error and the smallest interquartile range, which means that its prediction results are both accurate and stable. The long short-term memory model has a wider error distribution, and the standard recurrent neural network has the largest number of outliers. For an industrial robot, stable prediction is important because a single large error can trigger a false alarm or hide an actual fault trend.
The prediction branch is trained on periodic current segments. The training-to-testing ratio is 7:3, and the step length is 120. The predicted current signal is fed back to the TFO-BD diagnosis module. If the predicted signal matches a known fault pattern, the system outputs an early warning. This closed-loop design integrates prognosis and diagnosis and provides a practical path for predictive maintenance of an industrial robot.
7. Ablation and Mechanistic Discussion
To understand the contribution of each module, I analyze the roles of the time-domain filter, frequency-domain filter, and joint loss. The time-domain filter increases the kurtosis of the filtered signal, which means that impulsive fault components become more prominent. The frequency-domain filter increases the energy around the fault characteristic frequency, which improves the separability of different fault classes. The joint loss ensures that the blind deconvolution filters are not optimized only for a generic statistical objective but are also guided by the classification task. The mechanism is summarized in Table 9.
| Module | Primary effect | Physical interpretation | Expected benefit for industrial robot diagnosis |
|---|---|---|---|
| Time-domain quadratic filter | Suppresses non-periodic noise and amplifies impulsive components | Nonlinear modulation of fault impacts | Improves sensitivity to early bearing defects |
| Frequency-domain linear filter | Enhances fault characteristic frequency energy | Frequency response of the structural transmission path | Improves robustness to speed and load variation |
| Joint physical-informed loss | Aligns feature extraction with fault classification | Label information as a physical prior | Improves end-to-end fault discrimination |
| Gated recurrent unit prediction | Forecasts future current signal | Temporal evolution of the bearing state | Enables early warning and predictive maintenance |
The ablation logic can also be expressed through the optimization objective. Without the time-domain loss, the model loses the ability to enhance impulsive components. Without the frequency-domain loss, the model loses the ability to concentrate energy at the fault characteristic frequency. Without the cross-entropy loss, the filters may learn features that are statistically sparse but not discriminative for fault classification. The joint loss combines these objectives and allows the filters and the classifier to be optimized together. This is the main reason why TFO-BD improves multiple backbone networks under strong noise.
8. Engineering Implications and Limitations
The proposed method has several engineering implications for an industrial robot. First, the current signal can be collected from the motor drive without additional sensors, which reduces hardware cost and installation complexity. Second, the blind deconvolution filters are learned from data, so the method avoids manual bandpass parameter tuning. Third, the physical-informed loss provides a degree of interpretability, which is important when maintenance decisions must be explained to engineers. Fourth, the prediction-diagnosis closed loop supports early warning and can be integrated into a predictive maintenance system for an industrial robot.
However, this work has limitations. I only study single-component faults of the joint motor bearing. Other critical components, such as the reducer and the harmonic drive, are not included. In a real industrial robot, multiple faults may occur simultaneously, and the coupling between the bearing, reducer, and control system can make the current signal more complex. In addition, the operating conditions in this study are limited to two representative cases. More speed and load combinations should be tested to verify generalization. Future work will focus on multi-source information fusion and complex coupled fault diagnosis. I will also investigate noise-adaptive mechanisms and online learning strategies to improve the applicability of the method in real industrial robot systems.
9. Conclusion
In this work, I propose a time-frequency optimized blind deconvolution method for fault diagnosis and prognosis of an industrial robot joint motor bearing under strong noise. The method embeds a time-domain quadratic convolution filter and a frequency-domain linear filter into a deep learning framework. A physical-informed loss function combines kurtosis constraints, frequency-domain energy constraints, and cross-entropy classification loss. This design enables end-to-end joint optimization of the blind deconvolution filters and the fault classifier. Experimental results at \(-4\) dB show that the proposed method significantly improves multiple mainstream backbone networks, with an average F1 score increase of more than 10%. I also construct a gated recurrent unit multi-step prediction model that forecasts fault evolution and provides early warning feedback. The prediction and diagnosis modules form a closed-loop system for an industrial robot. The results confirm the effectiveness and engineering potential of the proposed method in strong-noise industrial scenarios.
