Below is an examiner-style critique and a set of concrete amendments aimed at improving patentability, clarity, support, and enforceability of the subject matter as claimed.
I. Principal deficiencies and risks under standard patentability criteria
1) Clarity/definiteness (support and enablement)
- The terms “edge-side,” “lightweight,” “anomaly score,” “key frequency band features,” and “periodically updated by federated learning without uploading raw data” are indefinite without operational definitions, parameter ranges, or objective boundaries. This creates vulnerability under clarity and enablement requirements (e.g., clarity and support provisions comparable to Article 26 of the CN Patent Law and 35 U.S.C. §112).
- The specification, as inferable from the claims, appears to omit critical implementation details: filter characteristics, window length/overlap, sampling rates, model architecture parameters, anomaly score calibration, and the federated protocol (client update rule, aggregation, security/privacy guarantees, and communication cadence). These omissions may undermine enablement and written description.
2) Novelty and inventive step
- Each of the following elements is widely disclosed in the art: tri-axial vibration sensing for rotating equipment, band-pass filtering and framing, 1D CNN-based anomaly scoring on edge devices, depthwise separable convolutions for resource-constrained inference, adaptive thresholds based on quantile estimation, and federated learning where raw time-series are retained on-device. Absent more specific technical interactions among these elements, the combination risks a finding of obviousness in view of predictive maintenance and edge analytics references.
- To improve inventive step, it is advisable to incorporate specific technical interactions and constraints tailored to rotating equipment (e.g., order-synchronous processing without a tachometer; frequency-band attribution tied to the learned filters; specific on-device resource envelopes; secure aggregation/privacy mechanisms that are compatible with non-IID vibration data). These features can generate synergistic effects and non-trivial advantages.
3) Subject-matter eligibility and unity
- The method subject matter is generally eligible. Unity of invention appears acceptable if federated updating is claimed as part of the same technical concept of on-device anomaly detection; however, the claims should make clear how the inference method and the federated update cooperate to achieve improved performance under edge constraints and privacy restrictions.
II. Recommended claim amendments
A. Amended independent method claim (replace Claim 1)
1. A computer-implemented method for detecting abnormal vibration in rotating equipment at an edge computing node, the method comprising:
(a) acquiring tri-axial vibration signals from a sensor rigidly affixed to a housing of a rotating machine at a sampling rate fs;
(b) pre-processing the signals by applying a band-pass filter having a passband [fL, fH] and linear-phase characteristics, and segmenting the filtered signals into frames of length W with overlap O;
(c) normalizing each frame by per-channel mean removal and variance scaling and generating a feature tensor;
(d) executing, on the edge computing node, an inference model comprising a lightweight one-dimensional convolutional neural network with depthwise separable convolutions and a final activation that maps to a bounded anomaly score S ∈ [0,1];
(e) determining an adaptive decision threshold τ by on-device quantile estimation computed over a sliding window of recent scores and optionally corrected for concept drift via exponential weighting;
(f) when S > τ, generating an alarm, storing a timestamp from a monotonic clock, and recording frequency-band attribution features computed by integrating gradient-based saliency over one or more key bands derived from the model’s learned filters; and
(g) periodically updating parameters of the inference model via a federated learning protocol in which the edge computing node computes local updates from on-device data and transmits only model parameters or clipped gradients to an aggregator implementing secure aggregation, without transmitting raw time-domain vibration signals.
B. Incorporate and refine Claim 2 and Claim 3 as follows
2. The method of claim 1, wherein the one-dimensional convolutional neural network comprises at least two depthwise separable convolutional blocks with residual connections, channel-wise attention, and post-training quantization to at most 8-bit precision to satisfy a memory footprint not exceeding M bytes and an inference latency not exceeding L milliseconds on the edge computing node.
3. The method of claim 1 or 2, wherein the adaptive threshold τ is determined as the q-quantile of the anomaly score distribution estimated by an exponentially weighted quantile estimator, wherein q ∈ [0.95, 0.999], and wherein τ is constrained by hysteresis or an N-of-M voting rule to reduce false alarms.
C. Add additional dependent claims to create fallback positions and strengthen inventiveness
4. The method of any preceding claim, wherein the band-pass filter passband [fL, fH] is set as a function of an estimated shaft rotational frequency fr obtained without a tachometer by detecting the dominant order in the spectrum and performing order-synchronous resampling.
5. The method of any preceding claim, wherein key frequency-band attribution features comprise integrated gradients or Grad-CAM-like attributions aggregated over K disjoint bands aligned to mechanical orders {1×, 2×, …}, and the system records top-p bands associated with the alarm to support root-cause analysis.
6. The method of any preceding claim, wherein federated learning employs FedProx or a proximal regularization to accommodate non-identically distributed (non-IID) vibration data across nodes and uses secure aggregation with randomized masking and differential privacy via per-update clipping and noise addition.
7. The method of any preceding claim, wherein the edge node rejects local updates if a concept-drift detector indicates instability of τ beyond a predefined bound, and defers to a server-supplied global model checkpoint.
8. The method of any preceding claim, further comprising performing envelope demodulation and computing spectral kurtosis within [fL, fH] as auxiliary inputs to the model.
9. The method of any preceding claim, wherein the edge node maintains a rolling buffer of only sufficient statistics, including quantile state, model gradients, and attribution summaries, to comply with a policy that prohibits storage or transmission of raw time-domain signals.
10. The method of any preceding claim, wherein the alarm is generated only upon satisfaction of both (i) S > τ and (ii) a persistence criterion over P consecutive frames, and the system logs a severity score proportional to the exceedance S − τ.
D. Add parallel independent claims to broaden coverage
System claim:
11. A system comprising: a tri-axial vibration sensor; an edge computing node with a processor and memory storing instructions which, when executed, perform the method according to any of claims 1–10; and a communication interface configured to transmit model parameters or gradients under a secure aggregation protocol without transmitting raw time-domain signals.
Computer-readable medium claim:
12. A non-transitory computer-readable medium storing instructions which, when executed by one or more processors of an edge computing node, cause the node to perform the method according to any of claims 1–10.
III. Specification enhancements to support the amended claims
1) Definitions and measurable constraints
- Define “edge computing node” (e.g., on-premise gateway or embedded microcontroller) and provide resource envelopes: CPU/DSP availability, memory budget (e.g., ≤1–4 MB), power, latency target (e.g., ≤50 ms/frame).
- Define “raw time-domain vibration signals” and make explicit what metadata, features, or model updates are permitted to be transmitted.
- Define “key frequency-band attribution features” and provide precise mathematical descriptions for saliency/attribution integration over frequency bands.
2) Signal acquisition and pre-processing
- Provide exemplary sampling rates (e.g., fs in 6–51.2 kHz), anti-aliasing, sensor mounting requirements, and tri-axial orientation calibration.
- Specify filter type (e.g., FIR linear-phase, order N, passband ripple/stopband attenuation) and frame/window parameters (W, O, window function).
- Include optional envelope demodulation, spectral kurtosis, and order tracking (tachometerless) methods.
3) Model architecture and optimization
- Describe the exact layer schema: number of depthwise separable conv blocks, kernel sizes, dilation factors, activation functions, residual/attention structures, and parameter count.
- Describe quantization, pruning, and knowledge distillation used to achieve “lightweight” constraints with numeric targets.
- Provide the anomaly score mapping (e.g., sigmoid) and training objective (e.g., one-class, self-supervised contrastive, or supervised cross-entropy), including loss functions.
4) Thresholding and concept drift
- Specify the quantile estimator (e.g., exponentially weighted quantile with update rate α), the quantile level q, hysteresis, and N-of-M rules.
- Include concept-drift detection criteria (e.g., KS distance on score distribution) and safe fallback behavior.
5) Federated learning protocol
- Describe local update computation (e.g., E epochs, batch size B), regularization (FedProx μ), and server aggregation (FedAvg or secure aggregation).
- Detail security/privacy: secure aggregation masks, key exchange, differential privacy parameters (clipping norm C, noise multiplier σ), and communication intervals.
- Provide resilience to non-IID data and intermittent connectivity (asynchronous updates, stale model handling).
6) Data governance and compliance
- Document that no raw time-domain data leave the edge; specify the retained logs (time stamps, attribution summaries, thresholds) and retention policy.
7) Performance evidence
- Provide experimental results on rotating machinery datasets: ROC/PR curves, latency/memory measurements on target hardware, robustness to non-stationarity, federated versus centralized baselines, and ablation of attribution and order-tracking features.
8) Alternative embodiments
- Variants with autoencoders, transformers with 1D convolutions, or hybrid statistical features as auxiliary inputs.
- Variants with on-device order-synchronous resampling tied to estimated shaft frequency.
IV. Drafting notes to minimize prosecution risk
- Avoid indefinite language such as “lightweight” or “key frequency band” without quantitative or algorithmic definition; anchor these terms to numeric bounds or procedural steps.
- Maintain explicit causal links between features and technical effects (e.g., how order-synchronous resampling and attribution features improve detection precision at constant bandwidth).
- Ensure claims do not preclude transmission of necessary training artifacts while still excluding raw waveforms; define “raw data” carefully.
- Include multiple fallbacks (e.g., with/without attribution; with/without DP; tachometerless order estimation optional) to preserve scope under close prior art.
V. Concluding assessment
With the above amendments and specification detail, the application is more likely to satisfy clarity and enablement, presents a clearer technical contribution tailored to edge-based vibration diagnostics, and offers non-trivial interactions among edge constraints, attribution-driven band logging, tachometerless order handling, and privacy-preserving federated updates. This should materially improve prospects against both novelty and inventive step challenges while preserving commercially meaningful scope.
