Time-frequency enhanced current diagnosis method

Abstract

The invention relates to the technical field of equipment health management and fault diagnosis, and particularly discloses a time-frequency enhanced current diagnosis method. According to the method, multi-phase current, temperature and rotating speed data are synchronously obtained, and a standardized sequence is generated through time alignment and noise suppression; self-adaptive spectrum analysis is carried out based on the working condition robust index, and a multi-scale energy diagram and frequency domain features are generated; fusion embedding is constructed through time domain feature extraction and time-frequency alignment mapping; and carrying out dynamic weight fusion and quality judgment in combination with the cross-modal consistency score, and finally generating a lightweight deployment model. According to the method, the anti-interference performance and interpretability of current diagnosis characteristics under multiple working conditions are effectively improved, accurate extraction and lightweight deployment of fault characteristics are realized, and the adaptability and reliability of a diagnosis system in a complex industrial environment are remarkably enhanced.

Description

Technical Field

[0001] This invention relates to the field of equipment health management and fault diagnosis technology, and in particular to a time-frequency enhanced current diagnosis method. Background Technology

[0002] For assessing the operational status of rotating equipment and electric drive systems, current signals, due to their ease of acquisition and broad coverage of fault mechanisms, have long been used for fault detection and location. Existing current diagnostic methods largely rely on time-frequency analysis with fixed parameters or single statistical descriptions, often resulting in characterization drift under operating conditions such as load fluctuations, speed changes, and gradual temperature rise. Furthermore, different fault modes exhibit significant differences in scale distribution across the time and frequency domains, making it difficult for a single characterization to simultaneously cover both early, weak anomalies and highly progressive anomalies, leading to sensitivity of identification boundaries and location results to operating conditions. To mitigate this issue, some methods attempt to introduce multimodal auxiliary quantities or multi-channel fusion, but these generally employ static weighting or rule-based thresholding, lacking data-driven channel-level and scale-level dynamic weight generation mechanisms. The consistency information across modes is not utilized in a structured manner, resulting in unclear boundaries between reliable and shielded regions during fusion, making it difficult to trace the judgment link back.

[0003] At the model implementation level, existing joint encoding methods mostly employ large-scale deep networks to establish mappings between time-domain and frequency-domain features in an end-to-end manner. However, their handling of input alignment, anomaly gating, and confidence management is rather coarse, lacking traceable organization at the segment and scale levels. In the threshold setting and result output stages, classification and localization are often completed using empirical boundaries or fixed confidence levels, without incorporating cross-modal consistency scores for threshold adaptation. This makes it difficult to stably produce structured initial judgment results and confidence levels under conditions of operating disturbances and multiple fault modes. Furthermore, existing closed-loop updates mostly rely on manual annotation or fully supervised retraining, resulting in a disconnect between training data organization and confidence management on the inference side. The construction of self-supervised sample pools lacks constraints from quality labels and prior weights, making it difficult to write back the judgment evidence from the inference side to the training side for continuous updates. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a time-frequency enhanced current diagnostic method, comprising: The system acquires multiphase current, temperature, and speed data, performs time alignment, noise suppression, robustness measurement and aggregation, and abnormal segment annotation processing to generate standardized sequences and robust operating condition indexes. Based on standardized sequences and robust indexes for operating conditions, parameter search, variable parameter analysis, resolution combination, and confidence assessment are performed to obtain multi-scale energy maps, frequency domain feature sets, and energy confidence masks. From the standardized sequence, multi-scale energy map, and energy confidence mask, the following steps are performed: extracting statistical features including amplitude distribution, peak-valley interval, and fluctuation density, and extracting temporal features including local fluctuations, rise and fall durations, etc.; performing time-frequency alignment operations using the abnormal segment index and the segment-level index of the standardized sequence as alignment references; and performing coordinate correspondence and scale matching operations using the energy confidence mask and the operating condition robust index as auxiliary information to generate a time-frequency alignment map. The system acquires a multi-scale energy map and a time-domain feature set, and performs joint encoding based on time-frequency alignment mapping and energy confidence mask, including shared trunk coding and branch coding including time-domain and frequency-domain branches. It also performs classification head and localization head reasoning processing based on cross-modal consistency score as a priori reference to obtain candidate categories, candidate positions, fusion embeddings and cross-modal consistency scores. Based on fusion embedding and cross-modal consistency scores, dynamic weight generation is performed, which uses cross-modal consistency scores as a reference for weight prior generation and feature contribution evaluation based on energy confidence masks. The dynamic weight entries are weighted and fused, and quality rule judgment processing including consistency verification and conflict resolution is performed to generate preliminary judgment results, confidence, quality labels and weight prior calibration information. Based on the initial judgment results and quality labels, sample screening, consistency training, and model compression operations are performed to build a lightweight deployment model.

[0005] Furthermore, the process of obtaining multiphase current, temperature, and rotational speed also includes: Multiphase current specifically refers to the phase current data collected by Hall effect sensors or current transformers installed on the phase conductors of each phase of the motor drive circuit. It includes the U, V, and W three-phase currents in a three-phase AC system and is used to characterize the real-time load and operating status of the equipment. Temperature specifically refers to thermal state data collected by temperature sensors attached to the heat dissipation surface of motor windings, bearing housings, or power devices. It includes one of the following: winding temperature, bearing temperature, or ambient temperature, and is used to monitor the thermal slow change process and thermal stress level of the equipment. Rotational speed specifically refers to the angular velocity data collected by a photoelectric encoder or magnetoelectric encoder installed on the motor shaft, which is used to characterize the real-time operating dynamics of the equipment.

[0006] Furthermore, the process of generating standardized sequences and robust operating condition indexes also includes: The system acquires multiphase current, temperature, and speed, performs time alignment with a unified timing signal within the system as a reference, and suppresses noise for power supply ripple, switching commutation, and environmental electromagnetic interference to obtain a standardized sequence containing phase and channel markers. Extract operating condition-related quantities related to load, speed change, thermal state, and power supply stability from the standardized sequence, perform robustness measures and time period aggregation based on fluctuation amplitude, duration, and cross-channel consistency, and obtain an operating condition robust index consistent with the standardized sequence segment index. The standardized sequence is subjected to abnormal segment screening and retention based on indicators such as intra-segment amplitude mutation, intra-segment morphological discontinuity, and cross-phase inconsistency. Segment index, channel marker, and timestamp marker are used to generate a correlation record between the abnormal segment index and the operating condition robust index.

[0007] Furthermore, the process of obtaining the multi-scale energy map, frequency domain feature set, and energy confidence mask also includes: Obtain standardized sequences and robust operating condition indexes, perform differentiated parameter searches for robust and non-robust time periods, and boundary constraints including time start and end markers and window length limits to obtain the configuration for adaptive spectrum analysis; The configuration for adaptive spectrum analysis is segmented according to the analysis window length and step interval in the configuration, and parametric analysis and resolution combination are performed according to the candidate frequency resolution set to generate a multi-scale energy map with time and scale as dual indices and a frequency domain feature set containing segment-level peak positions and bandwidth estimates. The reliability assessment of the multi-scale energy map and frequency domain feature set includes intra-segment consistency checks on energy continuity, relative order stability and synchronization, and threshold construction for inter-segment transition band processing. Energy confidence masks are generated and their mapping with the operating condition robust index is recorded.

[0008] Furthermore, the process of obtaining candidate categories, candidate positions, fused embeddings, and cross-modal consistency scores also includes: Statistical features, including amplitude distribution, peak-to-valley interval, and fluctuation density, as well as envelope correlation features, including local fluctuations and the duration of rise and fall, are extracted from the standardized sequence. Anomaly segment gating based on the anomaly segment index is then performed to obtain a time-domain feature set that corresponds one-to-one with the segment-level index of the standardized sequence. The time-domain feature set is aligned with the segment-level index of the abnormal fragment and the segment-level index of the normalized sequence. The time-domain feature set is then labeled with time periods and organized with a multi-level key index containing segment-level index, channel label and time position to obtain the time period mapping. The time-time mapping and multi-scale energy map are matched with coordinate correspondence and scale matching that matches the duration and envelope fluctuation of feature entries, using energy confidence mask and operating condition robust index as auxiliary information, to generate time-frequency aligned mapping.

[0009] Furthermore, the process of obtaining candidate categories, candidate positions, fused embeddings, and cross-modal consistency scores also includes: Obtain multi-scale energy maps and time-domain feature sets, and construct aligned inputs and generate masking entries and priority hints based on time-frequency alignment mapping and energy confidence mask to obtain inputs for time-frequency joint coding; The input used for time-frequency joint coding is subjected to shared trunk coding and branch coding including time-domain and frequency-domain branches to generate fusion embedding with segment level as the basic unit and cross-modal consistency scores of entry-level and segment-level records. The classification head and localization head are used for fusion embedding with cross-modal consistency score as a priori reference to generate candidate categories and candidate positions that correspond to the cross-modal consistency score.

[0010] Furthermore, the process of generating preliminary judgment results, confidence levels, quality labels, and weighted prior calibration information also includes: Obtain the fusion embedding and cross-modal consistency scores, use the cross-modal consistency scores as a reference to generate weight priors and evaluate the feature contribution based on energy confidence masks, and obtain dynamic weights that correspond one-to-one with the segment-level index; The dynamic weights and candidate categories and candidate positions are weighted and fused together, and a threshold is set based on source label and channel sorting to generate preliminary judgment results and confidence scores containing structured entries for category determination and position estimation. The initial judgment results and confidence levels are evaluated using quality rules that include consistency verification and conflict resolution, and the generated weight prior calibration information is labeled to produce quality labels and weight prior calibration information.

[0011] Furthermore, the process of building a lightweight deployment model also includes: The initial judgment results and quality labels are obtained. Based on the quality labels, the samples are screened and the structured organization, including the time-frequency alignment mapping coordinates and energy confidence mask list, is carried out to obtain a self-supervised sample pool with the segment level as the upper layer key and the channel as the middle layer key. Consistency training and parameter updates are performed on the self-supervised sample pool and weight prior calibration information, with the weight prior calibration information as the adjustment basis, to obtain the item-level and segment-level distilled attention and the channel-level accumulated sensitive channel list. Pruning is performed based on the list of sensitive channels, quantization is performed based on distilled attention, and distillation is performed based on teacher paths to generate a lightweight deployment model containing deployment metadata.

[0012] Furthermore, joint coding includes: Input units are read sequentially according to segment level, and a coding path with temporal and scale continuity is constructed by using the time coordinates and scale assignments recorded in the metadata. Upon receiving instructions from the shielding prompt, gating is implemented on areas marked as low confidence, so that the shared trunk only extracts common time-frequency structures around the effective area; By using channel labels as the grouping basis, a unified model is performed on the cross-channel correlation of multiphase currents to obtain the basic representation of cross-channels.

[0013] Furthermore, the branching code includes: In branch coding, the input is modally split according to the time-frequency alignment mapping: Within the time domain branch, local aggregation and cross-neighborhood interaction are performed around the temporal neighborhood of statistical features and envelope-related features, while retaining abnormal gating states and masking prompts to mark items that do not participate in the interaction. Within the frequency domain branch, local aggregation and cross-scale interaction are performed around the time-scale neighborhood of the energy patch, and low-confidence scale locations are suppressed based on the entries of the energy confidence mask. For cross-branch information exchange, one-to-one interaction is carried out based on the coordinate pairing table in the alignment mapping, so that time domain entries and frequency domain entries under the same time coordinate are paired and fused.

[0014] The key innovations of this invention include: (1) Segment-level index-driven standardization and robust operating condition modeling: output standardized sequences, robust operating condition indexes and abnormal segment indexes under the same time base, providing a constraint carrier for the boundary of spectral parameters and subsequent alignment.

[0015] (2) Frequency domain credibility management mechanism: Energy confidence mask is generated synchronously from multi-scale energy maps. Credible and shielded regions are expressed under time-scale dual indexes and run through subsequent gating and fusion.

[0016] (3) Traceable time-frequency alignment mapping: Time-domain features and frequency-domain representations establish a line-by-line correspondence between coordinates and scales, and input alignment and masking are accurately implemented at the item level.

[0017] The following are its main beneficial effects: (1) By denoising alignment and segment-level indexing, the standardized sequence, the robust index of operating conditions, and the index of abnormal segments are output on the same time base. The subsequent spectral parameter search and boundary constraints have a clear input range, reducing the representation drift caused by operating condition fluctuations.

[0018] (2) Adaptive spectrum analysis outputs multi-scale energy maps and frequency domain feature sets, and simultaneously generates energy confidence masks. The masks mark the credible region and the shielded region at the time and scale levels. Subsequent time domain feature extraction and alignment input construction obtain consistent screening criteria, reducing the probability of noise and transient disturbances entering the joint coding path.

[0019] (3) The temporal feature set and the multi-scale energy map achieve a line-by-line correspondence between coordinates and scales through time-frequency alignment mapping. In the joint encoding stage, the input is aligned and masked according to the segment level and channel label to reduce feature conflicts caused by cross-modal mismatch. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a time-frequency enhanced current diagnostic method provided in an embodiment of this application. Detailed Implementation

[0021] Example 1: Refer to Figure 1 This is a flowchart illustrating a time-frequency enhanced current diagnostic method provided in an embodiment of the present invention. The process may include at least steps S100-S600: S100: Acquire multiphase current, temperature and speed, perform time alignment, noise suppression, robustness measurement and aggregation, and abnormal segment annotation processing to generate standardized sequences and operating condition robust indexes; S200, based on standardized sequences and robust operating condition indexes, performs parameter search, variable parameter analysis, resolution combination, and confidence assessment to obtain multi-scale energy maps, frequency domain feature sets, and energy confidence masks. S300: From the standardized sequence, multi-scale energy map, and energy confidence mask, perform the following operations: extract statistical features including amplitude distribution characterization, peak-valley interval characterization, and fluctuation density characterization; extract temporal features including local fluctuations, rise and fall durations, and envelope correlation features; perform time period calibration using the abnormal segment index and the segment-level index of the standardized sequence as alignment references; and perform time-frequency alignment operations using the energy confidence mask and the operating condition robust index as auxiliary information to perform coordinate correspondence and scale matching, thereby generating a time-frequency alignment mapping. S400: Obtain multi-scale energy maps and time-domain feature sets, perform alignment input construction based on time-frequency alignment mapping and energy confidence mask, joint encoding of shared trunk coding and branch coding including time-domain and frequency-domain branches, and classification head and localization head reasoning processing based on cross-modal consistency scores as prior references to obtain candidate categories, candidate positions, fusion embeddings and cross-modal consistency scores; S500, based on fusion embedding and cross-modal consistency scores, performs dynamic weight generation by generating weight priors with cross-modal consistency scores as a reference and evaluating feature contribution based on energy confidence masks, weighted fusion of dynamic weight entries, and quality rule judgment processing including consistency verification and conflict resolution, generating preliminary judgment results, confidence levels, quality labels and weight prior calibration information; S600: From the initial judgment results and quality labels, perform sample screening, consistency training and model compression operations to build a lightweight deployment model.

[0022] Step S100 includes at least steps S110-S130: S110: Acquire multiphase current, temperature and rotation speed, perform time alignment and noise suppression to obtain a standardized sequence; In this step, the multiphase current specifically refers to the phase current data collected by Hall effect sensors or current transformers installed on the phase conductors of each phase of the motor drive circuit, including at least the U, V, and W phase currents in a three-phase AC system, used to characterize the real-time load and operating status of the equipment; the temperature specifically refers to the thermal state data collected by temperature sensors attached to the motor windings, bearing housings, or heat dissipation surfaces of power devices, including at least one of the winding temperature, bearing temperature, or ambient temperature, used to monitor the thermal slow-change process and thermal stress level of the equipment; the rotational speed specifically refers to the angular velocity data collected by photoelectric encoders or magnetoelectric encoders installed on the motor shaft, used to characterize the real-time operating dynamics of the equipment. The above multimodal data streams serve as the basic raw signals characterizing the overall electromechanical and thermal state of the equipment.

[0023] First, multiphase current, temperature, and speed data streams are synchronously acquired during continuous equipment operation. The multiphase current is generated by acquisition units installed in each phase circuit, while the temperature and speed are generated by corresponding sensing units. To ensure consistency in subsequent processing, data from different acquisition units are aligned according to a unified time reference, which uses a unified timing signal within the system as a reference. Time drift across channels and modes is eliminated by verifying the sampling start time and sampling interval. After alignment, noise suppression is applied to the multiphase current. This noise suppression, while ensuring that the boundaries of real events are not weakened, segments interference sources such as power supply ripple, switching commutation, and environmental electromagnetic interference. Temperature and speed are stabilized to reduce interference from sudden jumps in subsequent gating and mapping. Furthermore, amplitude and dimension unification are performed on the aligned and noise-suppressed data, mapping data with different dimensions to a unified numerical domain to ensure comparability and composability of cross-modal inputs in subsequent joint processing. After the above processing, a continuous and indexable standardized sequence is formed. This standardized sequence is segmented and identified according to time sequence, and phase and channel markers are retained in the data structure to facilitate subsequent coordinate mapping with the multi-scale energy map. To ensure a closed loop between the preceding and following steps, the standardized sequence is retained as input for subsequent steps. On the one hand, it is directly used when extracting operating condition-related quantities from the standardized sequence; on the other hand, it serves as a basic data source for configuration generation when obtaining the standardized sequence and the robust operating condition index for parameter search and boundary constraint determination. Simultaneously, it provides a consistent time axis and amplitude domain for extracting statistical features and envelope-related features from the standardized sequence. In summary, the input of this step is a data stream of multiphase current, temperature, and rotational speed, and the output is a standardized sequence, which will be continuously referenced and used in configuration generation and feature extraction in subsequent steps.

[0024] S120. Extract operating condition-related quantities from the standardized sequence, perform robustness measurement and aggregation, and obtain the operating condition robustness index. In this step, the standardized sequence is used as the sole input for processing. Specifically, operating condition-related quantities related to load, speed variation, thermal state, and power supply stability are extracted from the standardized sequence. These operating condition-related quantities include time-period indicators reflecting dynamic changes and persistence indicators reflecting thermal slow-change characteristics, and include segment-level markers for identifying operating phases. To ensure consistency in subsequent constraints, the operating condition-related quantities are represented as a set on a unified time axis. Interval alignment and inter-segment concatenation are performed on operating condition-related quantities from different sources to form a traceable set of time periods. Subsequently, robustness measures are performed on the set of time periods. These robustness measures are based on the fluctuation amplitude, duration, and cross-channel consistency of the operating condition-related quantities, combined with a weighted integration of tolerance to abnormal disturbances, to form a robust characterization for each time period. Further, the robustness characterizations of each time period are aggregated, maintaining consistency with the segment index of the standardized sequence during the aggregation process, so that the aggregation results can establish a one-to-one correspondence between the time period constraints and the subsequent configuration used for adaptive spectrum analysis. After measurement and aggregation, a robust operating condition index is generated. This index uses time periods as the basic unit and records robustness information related to load, speed, and thermal state. To ensure continuous use between steps, the robust operating condition index is marked as a time period measurement result originating from the same source as the standardized sequence. It will be used as a boundary constraint condition in subsequent parameter searches and boundary constraint calculations using the standardized sequence and the robust operating condition index. This constraint is used to limit the search space and resolution combination of adaptive spectral analysis, and also serves as reference information for subsequent credibility assessment and threshold construction of the multi-scale energy map. Furthermore, the robust operating condition index also serves as the base index for subsequently constructing associated records with the anomalous segment index, ensuring consistency between anomalous information and operating condition information in terms of time and segment-level identification. In summary, the input of this step is the standardized sequence, and the output is the robust operating condition index, which is used in subsequent steps for parameter constraints, credibility assessment, and associated record generation.

[0025] S130. Perform abnormal segment screening and annotation on the standardized sequence, and generate a record of association between the abnormal segment index and the operating condition robust index; In this step, the standardized sequence continues to be used as the basic data, combined with the obtained robust operating condition index for processing. Specifically, the standardized sequence is inspected segment by segment. First, while maintaining the time period division consistent with the robust operating condition index, the waveform behavior of multiphase current within a segment is jointly examined with the changes in temperature and speed within a segment. Suspicious time periods are initially screened based on indicators such as abrupt amplitude changes within a segment, discontinuities in the form within a segment, and inconsistencies across phases. The suspicious time periods obtained from the initial screening are further verified based on inter-segment continuity and cross-modal comparison relationships. Local disturbances that are difficult to attribute to changes in operating conditions are confirmed as candidate segments of anomalies. To ensure the traceability of candidate segments of anomalies, the segment index, channel marker, and timestamp marker consistent with the standardized sequence are retained during the confirmation process, and the synchronization relationship with temperature and speed is checked, so that candidate segments of anomalies can be directly located to the corresponding time coordinates and scale coordinates when performing coordinate correspondence and scale matching with the multi-scale energy map in the future. Subsequently, anomalous candidate segments are labeled, including the segment start and end positions, the channels involved, the correspondence with the normalized sequence, and the correspondence with the robust operating condition index. Based on the above labels, an anomalous segment index is generated. This index uses a segment-level structure consistent with the normalized sequence to ensure that anomalous segment gating can be directly performed when extracting statistical features and envelope correlation features from the normalized sequence. To achieve synchronous referencing of anomalous information and operating condition information, an association record of the robust operating condition index is established simultaneously with the generation of the anomalous segment index. This association record stores the correspondence between anomalous segments and adjacent robust time periods, so that when performing credibility assessment and threshold construction on the multi-scale energy map and the frequency domain feature set, the relative position of the anomalous and robust segments can be referenced to limit the threshold range. Furthermore, the associated records of the anomalous fragment index and the robust operating condition index serve as shared data across steps. When subsequently acquiring the multi-scale energy map and the temporal feature set for alignment input construction and masking, these records are used to mask low-confidence regions within anomalous fragments. They also provide segment-level basis for quality label generation during subsequent weighted fusion and threshold setting of the fusion embedding. To ensure closed-loop continuity, the anomalous fragment index is not only used for subsequent time-segment labeling and indexing of the temporal feature set, but also participates in the subsequent collation of records with the cross-modal consistency score, ensuring that candidate categories and positions maintain a dual association with both anomalous and robust characteristics at the segment level. In summary, the input to this step is the standardized sequence and the robust operating condition index, and the output is the associated records of the anomalous fragment index and the robust operating condition index. This output will be continuously used in subsequent parameter constraints, alignment input construction, masking, segment-level labeling, and quality label generation.

[0026] Understandably, the standardized sequence generated in S110 is directly used as the sole source for extracting operating condition-related quantities in S120, and the operating condition robust index generated in S120 is used in S130 to limit the time period boundaries for abnormal segment screening and labeling. Simultaneously, the association record between the abnormal segment index generated in S130 and the operating condition robust index provides a time period reference in the subsequent parameter search and boundary constraint of the standardized sequence and the operating condition robust index. It serves as a segment-level masking basis in the subsequent execution of variable parameter analysis and resolution combination from the configuration used for adaptive spectrum analysis. It acts as a gating condition when extracting statistical features and envelope-related features from the standardized sequence. It serves as a dual input for masking and alignment when constructing and masking the multi-scale energy map and the time-domain feature set. Ultimately, it permeates the input organization for weighted fusion and threshold setting of the dynamic weights, candidate categories, and candidate positions, forming a traceable link composed of the standardized sequence, the operating condition robust index, and the abnormal segment index.

[0027] The technical effects of steps S110 to S130 can be summarized as follows: By generating standardized sequences under a unified time reference and establishing robust operating condition indexes and abnormal segment indexes and their associated records within the same segment-level structure, subsequent adaptive spectrum analysis, time-domain feature gating, time-frequency aligned input construction and masking processing, weighted fusion and threshold setting have a consistent data foundation and segment-level reference, thereby maintaining the consistency of input organization and the traceability of segment-level identification throughout the entire process.

[0028] Step S200 includes at least steps S210-S230: S210. Obtain the standardized sequence and robust index of operating conditions, perform parameter search and boundary constraints, and obtain the configuration for adaptive spectrum analysis; In this step, the standardized sequence formed in the previous steps serves as the sole original data source, and the robust operating condition index formed in the previous steps serves as the constraint basis for configuration generation for adaptive spectrum analysis. Specifically, the standardized sequence is first traversed according to segment-level identifiers, reading the multiphase current channel data corresponding to each time period, as well as the temperature and speed data synchronized with that time period, and establishing a segment-level index table under the same time base. Subsequently, the robustness measurement results for each time period in the robust operating condition index are retrieved, distinguishing between robust and non-robust time periods. A wider parameter search range is set for robust time periods, while a convergent parameter search range is set for non-robust time periods, and the parameter search priority and sampling step strategy are registered for the two types of time periods respectively. Further, during the parameter search process, an analysis window length, step interval, and frequency resolution matching the dynamic changes of that time period are determined for each time period. The determination process follows the segment-level consistency principle, ensuring that the phase currents within the same time period adopt a consistent time domain segmentation scheme and establish a correspondence with the temperature and speed change rates corresponding to that time period. To ensure the effectiveness of subsequent resolution combinations, this step also generates a set of boundary constraints for each time period. This set includes start and end time markers consistent with the robust operating condition index, maximum and minimum window lengths to adapt to the rate of change, lower step limits to match noise interference, and intra-segment shielding hints to mask anomalies. Understandably, if the preceding steps have generated association records between the anomalous segment index and the robust operating condition index, sub-intervals within anomalous segments are identified during parameter search and boundary constraints to enable shielding strategies and record intra-segment weight hints in subsequent resolution combinations. Through the above processing, a configuration for adaptive spectrum analysis is output. This configuration uses time periods as basic units and includes the analysis window length, step interval, candidate frequency resolution set, and boundary constraint set corresponding to each time period, maintaining consistency with the segment-level index of the standardized sequence. This configuration serves as one of the sole inputs for the next step, is registered as a reference for subsequent confidence assessment and threshold construction, and is used in conjunction with the robust operating condition index in subsequent steps, forming a data association throughout the entire process of this invention.

[0029] S220. Perform parametric analysis and resolution combination from the configuration used for adaptive spectrum analysis to generate a multi-scale energy map and a set of frequency domain features; In this step, using the configuration for adaptive spectrum analysis as input, multiple rounds of parametric analysis are performed on the standardized sequence to generate energy characterizations corresponding to different time scales and frequency resolutions. Specifically, firstly, based on the analysis window length and step interval of each time period in the configuration, the multiphase current channel data within that time period is segmented and sliced, ensuring that the slice boundaries are consistent with the segment-level index and retaining synchronous references with temperature and rotational speed. Subsequently, according to the candidate frequency resolution set in the configuration, each slice is analyzed one by one to obtain the energy estimation results at the corresponding scale. To maintain cross-scale consistency, the energy results obtained at different resolutions are stacked hierarchically in a unified coordinate system under the same time coordinate, and the stacked results are spliced ​​in segment-level order to form a multi-scale energy characterization covering the entire segment. Further, channel alignment and amplitude normalization processing are performed on the energy characterizations of each phase current channel within the same time period to obtain energy patches that can be compared across channels, and the intra-segment change rates of temperature and rotational speed are recorded together for direct reference in subsequent cross-modal consistency evaluation and alignment input construction. After analyzing all time periods and candidate resolutions, the energy patches from each time period are combined in chronological and scale order to output a multi-scale energy map. This multi-scale energy map uses time and scale as dual indices to record the energy distribution of each phase channel at different scales. Simultaneously, based on the multi-scale energy map, a frequency domain feature set is extracted. This set includes segment-level peak positions and bandwidth estimates, relative intensity ranking between channels, and energy change trajectories related to temperature and rotational speed change rates. The correspondence between each feature and the source time period, source channel, and source scale is preserved. Understandably, if the configuration for adaptive spectrum analysis includes a masking indicator for anomalous segments, regions within these anomalous segments are marked when generating the multi-scale energy map, and a masking marker is added to the features of these regions in the frequency domain feature set. This facilitates differentiated processing during subsequent confidence assessment and threshold construction. After the above processing, this step outputs a multi-scale energy map and a frequency domain feature set. Both are consistent with the segment-level index of the standardized sequence and are associated with the segment-level records of the robust operating condition index. They serve as direct inputs for the next step of credibility assessment and threshold construction. They will also be called by subsequent steps that perform coordinate correspondence and scale matching from the time domain feature set to establish time-frequency alignment mapping and alignment input construction.

[0030] S230. Perform credibility assessment and threshold construction on the multi-scale energy map and frequency domain feature set, generate energy confidence mask and record the mapping with the working condition robust index; In this step, using multi-scale energy maps and frequency domain feature sets as inputs, and the robust index for operating conditions and the configuration for adaptive spectrum analysis as auxiliary information, the reliability assessment and threshold construction of energy information at different time periods and scales are completed. Specifically, firstly, according to the segment-level index, the robustness measure corresponding to the segment is read, and combined with the boundary constraint set registered in the configuration, the segment's multi-scale energy map is checked for intra-segment consistency, including the continuity of energy between adjacent scales, the relative order stability of multiphase channels under a consistent time coordinate, and the synchronicity with the intra-segment change rate of temperature and rotational speed; local areas that do not meet the consistency requirements are marked. Subsequently, based on the records of peak positions, bandwidth estimates, and relative intensities between channels in the frequency domain feature set, combined with the intra-segment consistency check results and anomaly shielding markers, intra-segment threshold recommendations are constructed. The threshold recommendations include value ranges and shielding prompts for different scales and channels. To ensure a smooth transition of thresholds over time, this step introduces a transition band processing between segments, connecting and buffering the threshold recommendations of adjacent time periods to form a continuous threshold sequence. Furthermore, an energy confidence mask is generated on the multi-scale energy map according to a threshold sequence, the energy confidence mask being correlated with the multi-scale energy... Figure 1 The system records data using both time and scale indices, clearly distinguishing between reliable regions and regions requiring weight reduction within each time period, and reserving independent mask entries for regions in anomalous segments. To achieve bidirectional referencing with the robustness index, this step establishes a mapping relationship between each mask entry and its source time period robustness record while generating the energy confidence mask, forming a mapping table with the robustness index. This table is used for subsequent construction of alignment inputs and masking processing for time-frequency joint coding. Understandably, the energy confidence mask outputs consistent with the segment-level indexes of the multi-scale energy map, frequency domain feature set, and normalized sequence, and includes segment-level identifiers, scale identifiers, channel identifiers, and threshold source identifiers in a structured manner. This allows it to be directly read from subsequent steps of coordinate correspondence and scale matching from the time domain feature set, used to establish time-frequency alignment mapping and implement masking strategies during alignment input construction. Simultaneously, this mask will also be referenced in subsequent steps of weight prior generation and feature contribution evaluation for the fusion embedding, providing sample-level and segment-level weight hints. Therefore, this step outputs a mapping table between the energy confidence mask and the operating condition robust index, and registers the above output, along with the multi-scale energy map and the frequency domain feature set, as shared inputs for subsequent processes, forming a complete closed loop from the standardized sequence, the operating condition robust index, the configuration for adaptive spectrum analysis to the multi-scale energy map, the frequency domain feature set, and the energy confidence mask.

[0031] In summary, steps S210 to S230, using the standardized sequence and the robust operating condition index as input, sequentially complete parameter search and boundary constraints, variable parameter analysis and resolution combination, confidence assessment and threshold construction, ultimately generating a multi-scale energy map, a frequency domain feature set, and an energy confidence mask. A mapping relationship is established with the robust operating condition index, achieving seamless integration between data and configuration, representation and evaluation. This ensures that subsequent time-domain feature gating, the establishment of time-frequency alignment mapping, the construction of inputs for time-frequency joint coding, and the prior references for dynamic weight fusion have a consistent segment-level foundation and reliable region identifiers, thus forming a traceable and verifiable closed-loop technical path at the process level.

[0032] Step S300 includes at least steps S310-S330: S310. Extract statistical features and envelope correlation features from the standardized sequence, perform outlier segment gating, and obtain a time-domain feature set; In this step, the standardized sequence is used as the sole original input, and processing is carried out under a time base and segment-level identifier consistent with the standardized sequence. Specifically, the standardized sequence is first read within segments to obtain continuous data segments of multiphase current and continuous data segments of temperature and rotational speed synchronized with it. Without changing the segment-level start and end positions and channel labels, each segment is mapped to a unified time index to ensure that the correspondence between cross-channels and cross-modes can be directly retrieved. After completing the intra-segment reading, the intra-segment waveform of the multiphase current is amplitude stabilized and drift corrected to keep the intra-segment baseline consistent and retain the time marks of abrupt change positions and zero-crossing positions; the intra-segment data of temperature and rotational speed are modeled for gradual variation to record the gradual variation trend for subsequent combination judgment with pulse behavior. Subsequently, statistical features and envelope-related features are extracted from the intra-segment time index of the standardized sequence. The statistical features include amplitude distribution characterization at the segment level, peak-valley interval characterization at the window level, and fluctuation density characterization at the intra-segment neighborhood level. The envelope-related features are obtained by estimating and tracking the intra-segment envelope of the multiphase current, recording the local fluctuations, rise and fall durations, and energy concentration areas of the envelope within the segment. To ensure that features are generated only within reliable time regions, this step introduces anomaly segment gating during feature extraction: using the anomaly segment index as the gating basis, time sub-intervals within anomaly segments are removed from the statistical and envelope calculation window, or marked as independent entries for subsequent masking or weighting. This gating is performed by comparing the segment-level start and end positions of the anomaly segment index, and the removal and retention records during the gating process are written into the intra-segment metadata to ensure that the relationship between each feature item and the anomaly segment can be traced later. Furthermore, to enhance cross-channel comparability, this step performs channel alignment and amplitude normalization on the statistical characteristics and envelope correlation characteristics of multiphase currents, retains channel and segment-level labels, and writes corresponding time indices for temperature and rotational speed into the metadata for consistency verification during subsequent construction of time-segment mapping and coordinate correspondence. Through the above processing, a time-domain feature set is formed, with segments as the basic unit and channels as the index. This time-domain feature set corresponds one-to-one with the segment-level index of the standardized sequence, and records gating status and anomaly correlation in the entries. This time-domain feature set serves as the direct input for the next step, used for time-segment calibration and index organization; simultaneously, it serves as the preliminary data source for subsequent alignment and masking processing of the multi-scale energy map and the time-domain feature set, providing a time-domain basis for the input alignment and masking strategy before subsequent joint encoding. The input of this step is the standardized sequence and the anomaly segment index; the output is the time-domain feature set and its intra-segment metadata related to anomaly gating. This output is continuously referenced in subsequent steps and interfaced with the multi-scale energy map.

[0033] S320. Perform time period labeling and indexing on the time domain feature set to obtain the time period mapping; In this step, the time-domain feature set is used as the sole feature input, and the anomalous segment index and the segment-level index of the standardized sequence are used as alignment references to complete the time period labeling and index organization. Specifically, firstly, under the segment-level index consistent with the standardized sequence, the time-domain feature set is expanded segment by segment. The statistical features and envelope-related features within each segment are precisely located according to their time position at which they were generated. Missing boundary time markers are supplemented, and feature splicing gaps across windows are repaired to ensure that the corresponding feature entry can be located at any time within the segment. Subsequently, using the anomalous segment index as the verification basis, the consistency between the start and end positions of all feature entries and the start and end positions of anomalous segments is verified: when a feature entry crosses the boundary of an anomalous segment, the entry is split into independent sub-entries, and the overlap relationship with the anomalous segment and the overlap relationship with non-anomalous time are recorded respectively; when a feature entry is completely within an anomalous segment, it is assigned an anomalous label, and a reference to the anomalous segment index is written into the entry, which facilitates the direct use of masking or weighting strategies when establishing coordinate correspondence and scale matching in the later stage. To ensure consistency within and outside the segment level, this step, after completing the intra-segment localization, uniformly indexes and organizes the feature entries of all segments, constructing a time-segment mapping structure with the segment-level index as the upper key, channel markers and time positions as the middle key, and feature type as the lower key. This time-segment mapping structure retains the original time coordinates, intra-segment relative time, anomaly identifiers, and gating states of the feature entries, and supplements cross-segment connection information, enabling direct retrieval of the corresponding time-domain feature entries by segment, channel, and time when docking with the multi-scale energy map. Furthermore, during the generation process, the time-segment mapping structure synchronously writes references to the operating condition robustness index, giving each segment's mapping entry operating condition robustness background information, thus providing a basis for adopting differentiated coordinate correspondence strategies at different robustness levels when generating time-frequency alignment mappings. After the above processing, a time-segment mapping consistent with the anomaly fragment index is output. The time-segment mapping maintains a one-to-one correspondence with the time-domain feature set and is completely consistent with the segment-level index of the standardized sequence in terms of data organization. This time-segment mapping will serve as one of the sole inputs for coordinate correspondence and scale matching in the next step. It will also be directly invoked later when aligning the multi-scale energy map with the temporal feature set and performing masking processing to generate inputs for time-frequency joint coding. The inputs for this step are the temporal feature set, the anomalous segment index, and the segment-level index of the normalized sequence. The output is the time-segment mapping, which is used in subsequent steps to establish a coordinate correspondence with the multi-scale energy map and associate it with the energy confidence mask.

[0034] S330. Perform coordinate correspondence and scale matching between the time period mapping and the multi-scale energy map to generate a time-frequency aligned mapping; In this step, the time period mapping is used as the sole input in the time domain, the multi-scale energy map is used as the sole input in the frequency domain, and the energy confidence mask and the robust operating condition index are used as auxiliary information to complete coordinate correspondence and scale matching, outputting a time-frequency aligned mapping. Specifically, firstly, under the segment-level index consistent with the normalized sequence, the segment-level identifier, channel label, and time position of each feature entry in the time period mapping are read, and the energy patch consistent with the segment-level identifier and the channel label is retrieved in the multi-scale energy map. While keeping the original time coordinates unchanged, the time position of the feature entry is mapped to the time coordinate axis of the energy patch, resulting in a one-to-one time coordinate pairing table. Subsequently, based on the scale index in the multi-scale energy map, each pair of time coordinates is scanned at different scales, selecting the scale layer that best matches the duration of the feature entry and the scale layer that best matches the envelope fluctuation of the feature entry. Based on the principle of intra-segment continuity, the scale selection is smoothed at adjacent time positions, so that similar feature entries have compatible scale assignments at adjacent positions. To ensure the reliability of the mapping, after completing the time coordinate pairing and scale assignment, this step calls the energy confidence mask to mark the scale positions within the low-confidence region. This marking is written into the mapping entry, and the entry records the corresponding mask source and its mapping relationship with the robust operating condition index. This allows subsequent inputs used for time-frequency joint coding to directly identify available and masked regions. Furthermore, for feature entries with anomaly markers in the time-time mapping, this step prioritizes referencing the start and end positions of the anomaly segment index when establishing coordinate correspondences. It establishes a consistent relationship between the corresponding time coordinates and scale positions and the masking entries in the mask, and writes this relationship into the masking prompt field of the mapping entry so that it directly takes effect in subsequent alignment input construction and masking processing. After completing the above processing, a time-frequency alignment mapping is generated. This time-frequency alignment mapping has a segment-level outer structure, a channel-level middle structure, and a time and scale-level inner structure. It records the coordinate correspondence and scale matching results of each time-domain feature entry and the multi-scale energy map, and retains the energy confidence mask association marker and the mapping reference to the robust operating condition index in the entry.The time-frequency alignment mapping, when output, maintains consistency with the segment-level indexes of the time-domain feature set, the multi-scale energy map, and the normalized sequence. It can be directly read by subsequent steps involving the alignment of the multi-scale energy map and the time-domain feature set for input construction and masking, and is used to construct the input for time-frequency joint coding. Simultaneously, the time-frequency alignment mapping serves as pre-processing information for sharing trunk coding and branch coding of the input for time-frequency joint coding, providing a coordinate-level and scale-level alignment basis for calculating cross-modal consistency scores and generating candidate categories and positions. Furthermore, it is invoked as a basis for determining segment-level reliable regions and masked regions when subsequently obtaining the fused embedding and the cross-modal consistency score for weight prior generation and feature contribution evaluation. In summary, under a unified time reference and segment-level identifier, S310 to S330, with the standardized sequence and the abnormal segment index as prerequisites, and with the multi-scale energy map and the energy confidence mask as the interface objects, sequentially complete the construction of the time-domain feature set, the formation of the time-segment mapping, the generation of the time-frequency alignment mapping, and the mask association, forming a traceable alignment link from the time domain to the frequency domain. This provides a directly usable mapping foundation for subsequent input construction and masking processing for time-frequency joint coding, the generation of fusion embedding and cross-modal consistency scores, and the prior reference of dynamic weights.

[0035] The technical effects of steps S310 to S330 can be summarized as follows: under the conditions of unified segment-level indexing, consistent anomaly gating, and clear scale matching, output time-frequency aligned mapping and complete the association of energy confidence masks, so that subsequent processes can perform input construction, representation learning, and weight fusion on a coherent data structure.

[0036] Step S400 includes at least steps S410-S430: S410. Obtain the multi-scale energy map and time-domain feature set, perform aligned input construction and masking processing to obtain the input for time-frequency joint coding; In this step, the multi-scale energy map output from the previous step is used as the frequency domain input, and the time domain feature set output from the previous step is used as the time domain input. Simultaneously, the time-frequency alignment mapping and energy confidence mask established in the previous step are used as the basis for alignment and masking, and a unified organization is performed around the segment-level index and channel labels. Specifically, firstly, under a time reference consistent with the standardized sequence, the time coordinates and scale assignments of each segment's time domain feature entry are read from the time-frequency alignment mapping, and energy blocks consistent with these time coordinates and scale assignments are retrieved from the multi-scale energy map. For boundary positions that cannot be directly aligned, neighborhood interpolation and intra-segment splicing are performed based on the adjacent coordinates recorded in the mapping, so that the aligned energy blocks and corresponding time domain feature entries form a one-to-one pairing input unit. Subsequently, around each input unit, the channel order and amplitude domain are unified: without changing the channel labels, the multiphase current channels are rearranged according to the agreed order; the amplitude characterization of the time domain features and the amplitude characterization of the frequency domain energy are mapped to a unified numerical range, and the segment-level parameters of this mapping are recorded for subsequent traceability. Furthermore, the input units are masked according to the energy confidence mask: masked entries are generated for time-scale regions marked as low confidence, these entries are associated with the corresponding time-domain feature entries, and masking and priority hints are written into the metadata of the input units; for time-domain feature entries containing anomaly identifiers, the corresponding masked entries in the energy confidence mask are read, so that these entries are uniformly identified and processed in subsequent encoding processes. After the above alignment and masking processes, all input units are assembled in segment-level and channel-level order to form the input for time-frequency joint coding; this input maintains index consistency with the multi-scale energy map, the time-domain feature set, and the time-frequency alignment mapping, and retains complete segment-level identifiers, channel markers, time coordinates, scale assignments, and masking hints in the metadata, thus serving as the sole data source for the next step of shared trunk coding and branch coding.

[0037] S420. The input used for time-frequency joint coding is subjected to shared trunk coding and branch coding to generate fusion embedding and cross-modal consistency scores; In this step, the input used for time-frequency joint coding is used as the sole input, and representation results and consistency quantization are generated around the collaborative process of shared trunk coding and branch coding. Specifically, firstly, in the shared trunk, input units are read sequentially according to segment-level order, and a coding path with temporal and scale continuity is constructed using the time coordinates and scale assignments recorded in the metadata. On this path, instructions from masking prompts are received, and gating is implemented on regions marked as low confidence, so that the shared trunk extracts common time-frequency structures only around effective regions. At the same time, channel labels are used as the grouping basis to uniformly model the cross-channel correlation of multiphase currents, obtaining the basic cross-channel representation. Subsequently, in the branch coding, the input is modally split according to the time-frequency alignment mapping: within the time domain branch, local aggregation and cross-neighborhood interaction are performed around the temporal neighborhood of statistical features and envelope-related features, retaining abnormal gating states and masking prompts to mark entries that do not participate in the interaction; within the frequency domain branch, local aggregation and cross-scale interaction are performed around the time-scale neighborhood of the energy patch, and low-confidence scale positions are suppressed according to the entries of the energy confidence mask; for cross-branch information exchange, one-to-one interaction is performed according to the coordinate pairing table in the alignment mapping, so that time domain entries and frequency domain entries under the same time coordinate are paired and fused. Furthermore, the common representations generated by the shared trunk and the modal representations produced by branch coding are concatenated and weighted according to the entry index to obtain a fusion embedding with the segment level as the basic unit. Simultaneously, based on the pairing relationship between each pair of time-domain and frequency-domain entries, aligned differential quantization and co-quantization are extracted, and the participation levels of masking hints and anomaly indicators are considered to calculate a cross-modal consistency score. This score is recorded at both the entry level and the segment level for direct reference in subsequent weight prior generation and candidate result organization. After the above processing, the fusion embedding and cross-modal consistency score are output. Both are consistent with the input used for time-frequency joint coding in terms of segment-level index and channel labeling, and carry alignment mapping references and mask references, serving as direct inputs for the next step of classification head and localization head inference.

[0038] S430. Perform classification head and localization head reasoning on the fused embedding to generate candidate categories and candidate positions; In this step, the fusion embedding is used as the unique input, and the cross-modal consistency score is used as a priori reference to organize and output the category determination and position estimation. Specifically, firstly, in the classification header, the set of fusion embedding entries is read sequentially according to the segment-level index. Entry-level aggregation is performed within each group based on the channel label, retaining the mask references and anomaly identifier references related to the aggregation, so that masked entries do not participate in aggregation or participate in a reduced-weighted manner. After aggregation, the aggregation results within the group and the segment-level aggregation results are structured to generate segment-level category candidate entries. A one-to-one correspondence value with the cross-modal consistency score is written into each entry to reflect the degree of cross-modal cooperation upon which the candidate entry was formed. Subsequently, in the positioning head, the set of fusion embedding entries corresponding to the classification head is read. Based on the time coordinates and scale assignments recorded in the time-frequency alignment mapping, candidate position entries are generated: adjacent entries are merged based on intra-segment continuity in the time coordinates, and adjacent assignments are smoothed based on compatibility in the scale assignments; entries involving masking prompts are downweighted or removed during the position generation stage, and the processing results are explicitly marked in the entries; the generated candidate position entries and candidate category entries are paired under the same segment-level index and channel label to form a one-to-one corresponding candidate category and candidate position. Further, in the candidate result organization stage, each pair of candidate categories and candidate positions is appended with its corresponding cross-modal consistency score and mask reference, and the segment-level reference to the robust operating condition index is retained, so that subsequent steps can differentiate between different segment-level reliable backgrounds based on dynamic weight fusion and threshold setting; simultaneously, the candidate results are written into a traceable record structure, recording the reference relationships with the fusion embedding, the alignment mapping, and the energy confidence mask, so that the next step can directly retrieve the source and constraints of the candidate entries at the sample-level and segment-level granularity. Through the above reasoning and organization, candidate categories and candidate positions are output, forming a correspondence table with the cross-modal consistency scores. This output serves as the direct input for subsequent weight prior generation and feature contribution evaluation of the fused embedding and the cross-modal consistency scores, and also as a reference for subsequent sample screening and structured organization. It is used to label the quality and consistency sources of entries when forming the self-supervised sample pool. Understandably, the input for time-frequency joint encoding formed in S410 is completely consumed and transformed into fused embedding and cross-modal consistency scores in S420. The fused embedding and cross-modal consistency scores formed in S420 are simultaneously referenced in S430 to generate candidate categories and candidate positions and establish correspondences. All three maintain consistency in segment-level indexing and channel labeling, thus forming a complete data loop with the multi-scale energy map, temporal feature set, time-frequency alignment mapping, and energy confidence mask output from previous steps.

[0039] The technical effects of steps S410 to S430 can be summarized as follows: By implementing shared trunk coding and branch coding based on the alignment input construction and masking processing, a fusion embedding and cross-modal consistency score carrying alignment references and mask references is generated. The correspondence between consistency and mask is simultaneously preserved in the organization of the classification head and the positioning head. Thus, candidate categories and candidate positions are output in a two-level structure of segment level and channel level, providing directly searchable alignment priors and reliable region identifiers for subsequent dynamic weight fusion and threshold setting.

[0040] Step 500 includes at least steps S510-S530: S510. Obtain the consistency score of fusion embedding and cross-modal operation, perform weight prior generation and feature contribution evaluation to obtain dynamic weights; In this step, the fusion embedding generated in the previous step is read first, and the cross-modal consistency score recorded in the previous step is read simultaneously. Metadata such as segment-level identifiers, channel tags, time coordinates, and scale assignments retained by the previous step are also retrieved to ensure consistent retrieval at the entry level. Specifically, for each segment's fusion embedding entry, a weighted prior is first established using the cross-modal consistency score as a reference. This weighted prior is grouped and registered within the segment according to the channel tag, and references to time-frequency alignment mapping and energy confidence masks are retained in the entry for differentiated processing in different trusted regions. Subsequently, the feature contribution of each time-domain and frequency-domain component in the fused embedding is evaluated. During the evaluation, the time coordinate and scale assignment of each entry are used as indexes. Masked entries and weighting hints are read from the energy confidence mask corresponding to that entry. Entries located in low-confidence regions are weighted, while entries passing the anomaly gate are retained. At the channel level, multiphase current entries are merged according to channel labels, and dominant and subordinate entries within a channel are registered. Subordinate entries are weighted during contribution aggregation to maintain consistent ordering across channels. Further, at the segment level, the results of weight priors and feature contribution evaluations are integrated item by item. Compatibility checks are performed on the same time neighborhood and adjacent scale neighborhoods, and duplicate entries are eliminated. A dynamic weighted entry set corresponding one-to-one with the segment-level index is output. This dynamic weighted entry set retains the channel label, time coordinate, scale assignment, and reference relationship with cross-modal consistency score and energy confidence mask for each entry, thus serving as direct input for the next step of weighted fusion and threshold setting. Simultaneously, to form a closed loop with subsequent self-supervised training, this step supplements the dynamic weight entry set with source markers. These markers indicate the entry's participation path in weight priors, feature contribution evaluation, and mask referencing, ensuring that the generation of subsequent quality rule determination and weight prior calibration information can be traced back to the entry's composition basis. The inputs to this step are the fusion embedding and cross-modal consistency scores, and the output is the dynamic weights. This output is fully utilized in subsequent weighted fusion and threshold setting, and is referenced again as an important reference in quality rule determination and labeling.

[0041] S520. Perform weighted fusion and threshold setting on dynamic weights, candidate categories and candidate positions to generate preliminary judgment results and confidence levels; In this step, the dynamic weights output from the previous step are read first, and the candidate categories and candidate positions formed in the previous step are read simultaneously. The two types of inputs are consistent in segment-level index and channel labeling, and share the retrieval method of time coordinate and scale assignment. Specifically, for each pair of candidate categories and candidate positions, dynamic weight entries in the same segment, channel, and time neighborhood are retrieved. Time coordinates and scale assignments that overlap with the pair of candidate entries are matched first, and the matched dynamic weights are used as the weighting coefficient source for the pair of candidate entries. When multiple dynamic weight entries partially overlap with the same pair of candidate entries, the dominant entries in the time neighborhood are prioritized, and the compatible entries in the scale neighborhood are secondary, and the processing traces of masked entries and weight reduction prompts during the integration process are retained. Subsequently, the candidate categories are weighted and aggregated, generating weighted results at both the channel and segment levels. The corresponding values ​​of the cross-modal consistency scores are written into the entries as parallel references. Candidate positions are then weighted and smoothed. Adjacent positions are merged based on intra-segment continuity on the time axis, and abrupt jumps are eliminated based on compatibility principles in scale assignment. Simultaneously, weighted segments within low-confidence regions of the energy confidence mask are either downweighted or removed. To ensure the output has executable decision boundaries, this step sets thresholds after weighting: at the segment level, differentiated thresholds are assigned to entries with different confidence backgrounds based on the source markers of the dynamic weight entries; at the channel level, relative thresholds are assigned to entries in different channels based on the order of dominant and subordinate entries within the channel; at the entry level, a more conservative threshold is set for entries containing masking hints, combining cross-modal consistency scores and mask references. After the above processing, the preliminary judgment result and confidence score are output. The preliminary judgment result includes structured entries for category determination and location estimation. The confidence score is recorded in the entries in a way that corresponds one-to-one with the segment-level index, channel label, time coordinate, and scale assignment. At the same time, the key parameters used when setting the threshold and their reference sources are written into the entry metadata so that subsequent quality rule determination can be based on this to conduct rule-based verification of entry quality. The input of this step is dynamic weights and candidate categories and candidate positions, and the output is the preliminary judgment result and confidence score. This output will serve as the direct input for the next step of quality rule determination and labeling, and as the basis for subsequent sample screening and structured organization in the construction process of the self-supervised sample pool.

[0042] S530. Perform quality rule judgment and labeling on the preliminary judgment results and confidence levels, and generate quality labels and weight prior calibration information. In this step, the initial judgment result and confidence level are read, and the key threshold setting parameters, reference sources, and correspondences with cross-modal consistency scores, energy confidence masks, and time-frequency alignment mappings recorded in the entry metadata of previous steps are read simultaneously to achieve entry-level traceability during rule-based verification. Specifically, firstly, within the segment level, the initial judgment result is verified for consistency using the key threshold setting parameters as a reference: when multiple pairs of candidate categories and candidate positions overlap for the same device part within the same segment, conflict resolution is performed according to the order of priority of dominant entries within the channel and cross-modal consistency scores, and the adoption order and reasons for elimination during the conflict resolution process are recorded in the entry metadata; when the same segment crosses the boundary with adjacent segments in location estimation, boundary convergence and merging are implemented with reference to the continuity prompts of the time-frequency alignment mapping. Subsequently, at the channel and item levels, based on the low-confidence regions and masking hints in the energy confidence mask, quality levels are assigned to corresponding items, and these quality levels are recorded as quality tags. For items containing traces of anomalous gating participation, gating source information is added to the quality tags to support differentiated selection in the subsequent self-supervised sample pool. Furthermore, weight prior calibration information is generated around each type of quality tag: when an item is downgraded due to a low-confidence region, the weight priors related to that item are adjusted downwards within the same segment and channel; when an item is downgraded due to insufficient cross-modal consistency scores, the weight priors related to that item are contracted within the same temporal neighborhood; when an item is eliminated due to conflict resolution caused by threshold setting boundaries, the weight priors related to that item are marked as candidate items not to be used in subsequent similar scenarios. The aforementioned weight prior calibration information is output in a two-tiered structure of segment-level and item-level, clearly indicating the adjustment direction and applicable scope to be adopted in the next weight prior generation and feature contribution evaluation. It also establishes a write-back association with the dynamic weight item set from the previous steps in the form of reference relationships and source tags. After rule-based judgment and labeling, quality labels and weight prior calibration information are output. The quality labels, along with the initial judgment results and confidence levels, enter the subsequent sample screening and structured processing flow. The weight prior calibration information enters the subsequent consistency training and parameter update flow, used for targeted convergence and preservation of the encoding path and weight strategy during the self-supervised training and lightweight distillation stages. Understandably, the dynamic weights generated by SX10 are retrieved one by one in SX20 and participate in weighted fusion and threshold setting. The initial judgment results and confidence levels generated by SX20 are verified one by one in this step and assigned quality labels. At the same time, writeable weight prior calibration information is formed. The three are consistent in segment-level index and channel label, and together with the fusion embedding, cross-modal consistency score, candidate category and candidate position, energy confidence mask and time-frequency alignment mapping in the previous steps, they form a complete data closed loop.

[0043] The technical effects of steps S510 to S530 can be summarized as follows: Under the constraints of unified segment-level index and channel labeling, through weight prior generation and feature contribution evaluation, weighted fusion and threshold setting, as well as quality rule judgment and labeling, a continuous output link is formed from dynamic weights to preliminary judgment results and confidence levels, and then to quality labels and weight prior calibration information, providing a write-back and traceable basis for subsequent sample screening, consistency training and parameter updates.

[0044] Step S600 includes at least steps S610-S630: S610. Obtain preliminary judgment results and quality labels, perform sample screening and structured organization, and obtain a self-supervised sample pool. In this step, the preliminary judgment result and quality label generated in the previous step are used as the sole judgment criteria. Simultaneously, the segment-level index, channel label, time coordinate, scale assignment, and reference information associated with the preliminary judgment result, as well as the cross-modal consistency score, energy confidence mask, and time-frequency alignment mapping, are retrieved to ensure that the sample organization remains consistent with the data interface for subsequent training. Specifically, the preliminary judgment result is first traversed at the segment level, reading the category and position information corresponding to each entry, and simultaneously retrieving the quality label of the same segment and channel. When the quality label indicates that an entry can be used for self-supervised learning, the entry is recorded as a candidate self-supervised sample; when the quality label indicates that an entry is affected by masking or weighting prompts, the entry is retained but a weak supervision label is added, serving as a sample source under different weighting conditions. Subsequently, within the candidate self-supervised sample set, a priority sequence of samples is established based on the correspondence with the cross-modal consistency score, and time-scale segments located in the low-confidence region of the energy confidence mask are eliminated or weighted, ensuring that retained and eliminated segments have distinguishable identifiers within the same entry. To ensure data consistency with subsequent consistency training, this step further generates a structured record for each candidate self-supervised sample. This structured record includes: corresponding coordinates of time-domain and frequency-domain segments consistent with the time-frequency alignment mapping; a list of trusted and masked regions consistent with the energy confidence mask; placeholder fields for subsequent docking with the weight prior calibration information; and the original segment-level index and channel label mapped to the normalized sequence. After traversing and organizing all entries, the sample proportion of each segment and channel is controlled according to a segment-level and channel-level balancing strategy to avoid excessive concentration of samples in a single operating condition or channel, ultimately outputting a self-supervised sample pool. This self-supervised sample pool uses segments as the upper-level key, channels as the middle-level key, and structured records as the entry unit, serving as the sole data input for the next step of consistency training and parameter updates. A reference relationship with the initial judgment result and the quality label is written in the data header for subsequent traceability.

[0045] S620. Perform consistent training and parameter updates on the self-supervised sample pool and weight prior calibration information to obtain the distillation attention and sensitive channel list. In this step, the self-supervised sample pool is used as the training data source, and the weight prior calibration information is used as the adjustment basis. Simultaneously, the time-frequency alignment mapping and the energy confidence mask bound to the sample entries are read, allowing the training process to unfold under the constraints of alignment coordinates and reliable regions. Specifically, at the entry level, the self-supervised sample pool is first divided into a strong sample set and a weak sample set: the strong sample set refers to entries whose structured records have high reliable region coverage and good consistency with the quality label; the weak sample set refers to entries whose records contain masked regions or reduced-weight segments and have shrinkage prompts with the weight prior calibration information. In terms of training organization, the strong sample set is directly read according to the time-frequency alignment mapping, corresponding to time-domain segments and frequency-domain segments, maintaining a one-to-one correspondence between segments on the time-scale coordinates; for the weak sample set, low-reliability regions marked by the energy confidence mask are skipped during reading, and the skipped segments are registered as segments that do not participate in consistency constraints, to avoid introducing unstable factors in parameter updates. Subsequently, the weight prior calibration information is invoked to differentiate the participation intensity of sample items from different sources during training: when an item is downgraded in a previous step due to insufficient cross-modal consistency score, a contraction intensity is assigned to the item in this step to reduce its influence on common parameters; when an item is downgraded because it is located in a low-confidence region, a down-adjustment intensity is assigned to the item in this step, so that it only participates in constraints on unmasked and trusted segments. To generate attention information that can be referenced for subsequent lightweighting and decision fusion, this step counts key positions that are sampled and retained multiple times in the interaction between time-domain and frequency-domain segments on the parallel path of consistency training, forming item-level and segment-level attention accumulation records; and at the channel dimension, channel markers that are repeatedly retained in different segment levels and contribute significantly to constraint convergence are counted, forming channel-level accumulation records. After multiple rounds of parameter updates and statistical accumulation, this step integrates the accumulated records at the item and segment levels into distilled attention to reflect the distribution of key positions under alignment coordinates and confidence region constraints. Simultaneously, the channel-level accumulated records are organized into a sensitive channel list, identifying the set of channels that have consistently participated in and are effective against constraints across multiple data segments and samples. The output distilled attention and sensitive channel list share segment-level indices and channel labels with the self-supervised sample pool, and establish a write-back reference with the weight prior calibration information, allowing them to be directly retrieved and applied in the model configuration and deployment process in the next step.

[0046] S630 performs pruning, quantization, and distillation to generate a lightweight deployment model; In this step, the model structure and parameter snapshot consistent with the previous training are first loaded. The distilled attention and the sensitive channel list are read as lightweight constraints. The organization of the data to be processed is configured with an input reduction matching the normalized sequence to ensure that the requirements for input format and indexing system in the pruning, quantization, and distillation stages are consistent with those in the training stage. Specifically, in the pruning stage, the model channel dimensions are processed hierarchically according to the sensitive channel list: the channels included in the sensitive channel list maintain structural integrity, and their connections are only moderately shrunk in the low-contribution regions corresponding to the energy confidence mask; the channels not included are structurally reduced according to the long-term contribution of segment-level statistics, and the alignment operator corresponding to the time-frequency alignment mapping is retained during the reduction to ensure that the model can still receive time-domain and frequency-domain segments used for alignment input construction after the reduction. During the quantization phase, the distilled attention is read, and operators corresponding to high-value attention positions are placed in higher-precision quantization groups, while operators corresponding to low-value attention positions are placed in lower-precision quantization groups. A triple mapping of segment-level index, channel label, and scale assignment is written into the quantization table to ensure that numerical changes caused by quantization are traceable. Subsequently, during the distillation phase, the teacher path obtained from consistency training is invoked, and input batches isomorphic to the self-supervised sample pool are used to perform parameter correction on the lightweight network: at the entry level, key positions are aligned based on the distilled attention; at the channel level, the channel diffusion range of distilled energy is limited according to the sensitive channel list to prevent perturbations of non-critical channels from being introduced into the main path of the lightweight model. To create a deployable output, this step encapsulates the pruned, quantized, and distilled parameters and runtime configuration into a lightweight deployment model. The associated records with the distilled attention and the sensitive channel list are written into the deployment metadata. These records explicitly mark the set of channels preserved or protected in the model, the set of sub-operators assigned high precision, and the alignment strategies corresponding to key positions. This ensures that subsequent online inference maintains a consistent data path and interpretable weight path when referencing the energy confidence mask, the cross-modal consistency score, and the weight prior calibration information. Through this process, a lightweight deployment model is output, and a one-to-one reference mapping is established between it and the distilled attention and the sensitive channel list. This lightweight deployment model receives data organization isomorphic to the standardized sequence in subsequent online stages, forming a closed loop with the preceding steps. It can be used to generate new preliminary judgment results and quality markers for continuous updates to the self-supervised sample pool and continuous parameter correction.

Claims

1. A time-frequency enhanced current diagnostic method, characterized in that, include: The system acquires multiphase current, temperature, and speed data, performs time alignment, noise suppression, robustness measurement and aggregation, and abnormal segment annotation processing to generate standardized sequences and robust operating condition indexes. Based on standardized sequences and robust indexes for operating conditions, parameter search, variable parameter analysis, resolution combination, and confidence assessment are performed to obtain multi-scale energy maps, frequency domain feature sets, and energy confidence masks. From the standardized sequence, multi-scale energy map, and energy confidence mask, the following steps are performed: extracting statistical features including amplitude distribution, peak-valley interval, and fluctuation density, and extracting temporal features including local fluctuations, rise and fall durations, etc.; performing time-frequency alignment operations using the abnormal segment index and the segment-level index of the standardized sequence as alignment references; and performing coordinate correspondence and scale matching operations using the energy confidence mask and the operating condition robust index as auxiliary information to generate a time-frequency alignment map. The system acquires a multi-scale energy map and a time-domain feature set, and performs joint encoding based on time-frequency alignment mapping and energy confidence mask, including shared trunk coding and branch coding including time-domain and frequency-domain branches. It also performs classification head and localization head reasoning processing based on cross-modal consistency score as a priori reference to obtain candidate categories, candidate positions, fusion embeddings and cross-modal consistency scores. Based on fusion embedding and cross-modal consistency scores, dynamic weight generation is performed, which uses cross-modal consistency scores as a reference for weight prior generation and feature contribution evaluation based on energy confidence masks. The dynamic weight entries are weighted and fused, and quality rule judgment processing including consistency verification and conflict resolution is performed to generate preliminary judgment results, confidence, quality labels and weight prior calibration information. Based on the initial judgment results and quality labels, sample screening, consistency training, and model compression operations are performed to build a lightweight deployment model.

2. The method according to claim 1, characterized in that, The process of obtaining multiphase current, temperature, and speed also includes: Multiphase current specifically refers to the phase current data collected by Hall effect sensors or current transformers installed on the phase conductors of each phase of the motor drive circuit. It includes the U, V, and W three-phase currents in a three-phase AC system and is used to characterize the real-time load and operating status of the equipment. Temperature specifically refers to thermal state data collected by temperature sensors attached to the heat dissipation surface of motor windings, bearing housings, or power devices. It includes one of the following: winding temperature, bearing temperature, or ambient temperature, and is used to monitor the thermal slow change process and thermal stress level of the equipment. Rotational speed specifically refers to the angular velocity data collected by a photoelectric encoder or magnetoelectric encoder installed on the motor shaft, which is used to characterize the real-time operating dynamics of the equipment.

3. The method according to claim 1, characterized in that, The process of generating standardized sequences and robust operating condition indexes also includes: The system acquires multiphase current, temperature, and speed, performs time alignment with a unified timing signal within the system as a reference, and suppresses noise for power supply ripple, switching commutation, and environmental electromagnetic interference to obtain a standardized sequence containing phase and channel markers. Extract operating condition-related quantities related to load, speed change, thermal state, and power supply stability from the standardized sequence, perform robustness measures and time period aggregation based on fluctuation amplitude, duration, and cross-channel consistency, and obtain an operating condition robust index consistent with the standardized sequence segment index. The standardized sequence is subjected to abnormal segment screening and retention based on indicators such as intra-segment amplitude mutation, intra-segment morphological discontinuity, and cross-phase inconsistency. Segment index, channel marker, and timestamp marker are used to generate a correlation record between the abnormal segment index and the operating condition robust index.

4. The method according to claim 1, characterized in that, The process of obtaining the multi-scale energy map, frequency domain feature set, and energy confidence mask also includes: Obtain standardized sequences and robust operating condition indexes, perform differentiated parameter searches for robust and non-robust time periods, and boundary constraints including time start and end markers and window length limits to obtain the configuration for adaptive spectrum analysis; The configuration for adaptive spectrum analysis is segmented according to the analysis window length and step interval in the configuration, and parametric analysis and resolution combination are performed according to the candidate frequency resolution set to generate a multi-scale energy map with time and scale as dual indices and a frequency domain feature set containing segment-level peak positions and bandwidth estimates. The reliability assessment of the multi-scale energy map and frequency domain feature set includes intra-segment consistency checks on energy continuity, relative order stability and synchronization, and threshold construction for inter-segment transition band processing. Energy confidence masks are generated and their mapping with the operating condition robust index is recorded.

5. The method according to claim 1, characterized in that, The process of obtaining candidate categories, candidate positions, fused embeddings, and cross-modal consistency scores also includes: Statistical features, including amplitude distribution, peak-to-valley interval, and fluctuation density, as well as envelope correlation features, including local fluctuations and the duration of rise and fall, are extracted from the standardized sequence. Anomaly segment gating based on the anomaly segment index is then performed to obtain a time-domain feature set that corresponds one-to-one with the segment-level index of the standardized sequence. The time-domain feature set is aligned with the segment-level index of the abnormal fragment and the segment-level index of the normalized sequence. The time-domain feature set is then labeled with time periods and organized with a multi-level key index containing segment-level index, channel label and time position to obtain the time period mapping. The time-time mapping and multi-scale energy map are matched with coordinate correspondence and scale matching that matches the duration and envelope fluctuation of feature entries, using energy confidence mask and operating condition robust index as auxiliary information, to generate time-frequency aligned mapping.

6. The method according to claim 1, characterized in that, The process of obtaining candidate categories, candidate positions, fused embeddings, and cross-modal consistency scores also includes: Obtain multi-scale energy maps and time-domain feature sets, and construct aligned inputs and generate masking entries and priority hints based on time-frequency alignment mapping and energy confidence mask to obtain inputs for time-frequency joint coding; The input used for time-frequency joint coding is subjected to shared trunk coding and branch coding including time-domain and frequency-domain branches to generate fusion embedding with segment level as the basic unit and cross-modal consistency scores of entry-level and segment-level records. The classification head and localization head are used for fusion embedding with cross-modal consistency score as a priori reference to generate candidate categories and candidate positions that correspond to the cross-modal consistency score.

7. The method according to claim 1, characterized in that, The process of generating preliminary judgment results, confidence levels, quality labels, and weighted prior calibration information also includes: Obtain the fusion embedding and cross-modal consistency scores, use the cross-modal consistency scores as a reference to generate weight priors and evaluate the feature contribution based on energy confidence masks, and obtain dynamic weights that correspond one-to-one with the segment-level index; The dynamic weights and candidate categories and candidate positions are weighted and fused together, and a threshold is set based on source label and channel sorting to generate preliminary judgment results and confidence levels containing structured entries for category determination and position estimation. The initial judgment results and confidence levels are evaluated using quality rules that include consistency verification and conflict resolution, and the generated weight prior calibration information is labeled to produce quality labels and weight prior calibration information.

8. The method according to claim 1, characterized in that, The process of building a lightweight deployment model also includes: The initial judgment results and quality labels are obtained. Based on the quality labels, the samples are screened and the structured organization, including the time-frequency alignment mapping coordinates and energy confidence mask list, is carried out to obtain a self-supervised sample pool with the segment level as the upper layer key and the channel as the middle layer key. Consistency training and parameter updates are performed on the self-supervised sample pool and weight prior calibration information, with the weight prior calibration information as the adjustment basis, to obtain the item-level and segment-level distilled attention and the channel-level accumulated sensitive channel list. Pruning is performed based on the list of sensitive channels, quantization is performed based on distilled attention, and distillation is performed based on teacher paths to generate a lightweight deployment model containing deployment metadata.

9. The method according to claim 1 or 6, characterized in that, Joint coding includes: Input units are read sequentially according to segment level, and a coding path with temporal and scale continuity is constructed by using the time coordinates and scale assignments recorded in the metadata. Upon receiving instructions from the shielding prompt, gating is implemented on areas marked as low confidence, so that the shared trunk only extracts common time-frequency structures around the effective area; By using channel labels as the grouping basis, a unified model is performed on the cross-channel correlation of multiphase currents to obtain the basic representation of cross-channels.

10. The method according to claim 1 or 6, characterized in that, Branch coding includes: In branch coding, the input is modally split according to the time-frequency alignment mapping: Within the time domain branch, local aggregation and cross-neighborhood interaction are performed around the temporal neighborhood of statistical features and envelope-related features, while retaining abnormal gating states and masking prompts to mark items that do not participate in the interaction. Within the frequency domain branch, local aggregation and cross-scale interaction are performed around the time-scale neighborhood of the energy patch, and low-confidence scale locations are suppressed based on the entries of the energy confidence mask. For cross-branch information exchange, one-to-one interaction is carried out based on the coordinate pairing table in the alignment mapping, so that time domain entries and frequency domain entries under the same time coordinate are paired and fused.

Images