Abstract: To address latent faults in substation secondary protection systems, including protection setting drift, sampling link abnormalities, communication disturbances, and circuit breaker control loop failures, this paper proposes an edge-cloud collaborative fault diagnosis and hierarchical fault-tolerant disposal method. First, a lightweight transient event detection network, TED-Net, is deployed at the substation edge to screen steady-state and abnormal segments, thereby reducing full- volume data uploading. Second, wavelet packet decomposition and Gramian angular summation field mapping are used to construct multi-domain time-frequency feature matrices, and a cross-scale multi-domain feature fusion network, CS-MDFN, is developed to identify three anomaly levels: hidden drift, performance degradation, and rigid failure. Finally, the diagnosis levels are mapped to parameter correction, protection logic reconfiguration, and redundant module takeover, with setting- value verification, safety blocking, and concurrent-fault priority rules introduced. real time digital simulator simulation results show that TED-Net achieves an edge inference time of 7.2 ms and a transient anomaly recall rate of 98.4%; CS-MDFN reaches a comprehensive diagnostic accuracy of 97.1%; and the closed-loop disposal time in a typical rigid failure scenario is 75.6 ms. The proposed method can reduce communication and computing pressure while improving anomaly identification and hierarchical disposal capability for secondary protection systems.