In coal mine non-solidly grounded systems, high-impedance faults generate minimal zero-sequence currents with obscured characteristics and strong interference, complicating faulted line identification. Existing methods rarely address three-phase imbalance and variable cable parameters, causing selection errors. To this end, a method for identifying the non-effective ground fault routing of mining cables based on Steady-State Impedance Analysis (SSIA) and Holmes–Duffing oscillator small-signal detection is proposed. Firstly, based on SSIA, the mapping relationship that the phase of the zero-sequence current variation in the faulted line is the same as the phase of its voltage relative to the faulted ground is derived before and after the occurrence of the fault. Meanwhile, identifiable differences exist in both phase and amplitude of the zero-sequence current change in faulty lines compared to non-faulty lines before and after fault occurrence. This is used as the criterion for high-impedance ground fault line selection. In the mining environment, zero-sequence current variations are characterized as weak signals, which poses significant challenges for detection. Thus, a Holmes–Duffing oscillator weak signal detection method is proposed. Based on chaotic principles, accurate line selection is achieved by diagnosing chaotic states in oscillator-generated phase trajectories. A specific mine grid simulation via MATLAB/Simulink 2023b validates the method’s efficacy and applicability.
To enhance the level of automation in reading both pointer-type and mechanical dial-type instruments while mitigating cumulative measurement errors, this study proposes an instrument reading recognition algorithm built upon the YOLOv8n framework. A dual-stream processing architecture is employed to ensure accurate and robust meter readings across both instrument types. For pointer-type meters, the YOLOv8n-Pose network is utilized to detect keypoint coordinates corresponding to the dial center and needle tip. Meter readings are accurately determined by integrating geometric angle calculations with a two-point linear interpolation strategy. In addition, the YOLOv8n-Detect model is applied to localize specific numerical targets on mechanical dial-type meters. Following spatial ordering and region-of-interest (ROI) cropping, the extracted targets are subsequently input into a ResNet18 convolutional neural network for character classification. Experimental results demonstrate that the proposed system maintains high detection accuracy and reading stability under diverse illumination conditions and background interference, thereby providing a reliable technical solution for automated industrial inspection.