A deep learning-based permanent magnet motor fault diagnosis method

Abstract

The present application relates to the technical field of permanent magnet motor fault diagnosis, and particularly relates to a permanent magnet motor fault diagnosis method based on deep learning. The method comprises: based on rotor surface temperature values and stator coil temperature values, field collection and channel corresponding processing are performed to obtain a working condition data set; based on the working condition data set, bearing vibration frequency arrangement, bearing rotating speed arrangement and audio signal corresponding processing are performed to generate a mechanical screening result; then, fixed time window size segmentation, time synchronization, fault label cross comparison and comprehensive evaluation processing are performed on the mechanical screening result to generate a comprehensive evaluation result; and based on the comprehensive evaluation result, alarm signal generation and maintenance suggestion generation processing are performed to obtain a fault diagnosis result. The present application can straighten out the record relationship under the time window, and improve the continuity and reviewability of the fault diagnosis result generation process.

Description

Technical Field

[0001] This invention relates to the field of permanent magnet motor fault diagnosis technology, and in particular to a method for diagnosing permanent magnet motor faults based on deep learning. Background Technology

[0002] In the field of permanent magnet motor fault diagnosis technology, existing solutions typically involve collecting, processing, analyzing, and determining rotor surface temperature, stator coil temperature, bearing vibration frequency, bearing speed, and audio signals. These solutions suffer from limitations such as unclear connections between time-window segmentation results and time-synchronized processing, insufficient correlation between synchronous screening results and fault tag cross-comparison, and unclear sources of comprehensive evaluation results. Existing methods often employ a processing path of first processing the collected data and then directly determining the fault or outputting an alarm. Under fixed time-window segmentation, this easily leads to the mixing of records from different sources and unclear determination relationships within the same time window. This makes it difficult to reliably generate alarm signals and maintenance suggestions based on comprehensive evaluation results, ultimately leading to fault diagnosis results.

[0003] Existing technologies for the joint processing of operating condition datasets, mechanical screening results, time window segmentation results, synchronous screening results, and comprehensive evaluation results generally suffer from common shortcomings, such as incomplete time synchronization processing, lack of unified basis for cross-comparison of fault labels, and difficulty in preserving the relationship of previous records in comprehensive evaluation processing. These shortcomings make it difficult to form a consistent process in permanent magnet motor fault diagnosis scenarios, including field collection and channel correspondence processing, fixed time window segmentation processing, time synchronization processing, cross-comparison of fault labels and comprehensive evaluation processing, alarm signal generation and maintenance suggestion generation. This results in unclear record correspondence and discontinuous processing links in the fault diagnosis result generation process, and causes inconvenience to equipment operation monitoring, fault handling arrangements, and subsequent maintenance operations. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a deep learning-based method for diagnosing faults in permanent magnet motors, comprising:

[0005] S100. Based on the rotor surface temperature value and the stator coil temperature value, perform field aggregation and channel correspondence processing to obtain the operating condition dataset; the rotor surface temperature value refers to the temperature acquisition value of the outer surface of the permanent magnet motor rotor at the corresponding sampling time under the operating state; the stator coil temperature value refers to the temperature acquisition value of each acquisition point of the stator coil at the corresponding sampling time under the operating state.

[0006] S200. Based on the aforementioned working condition dataset, the bearing vibration frequency is processed, the bearing speed is processed, and the audio signal is correspondingly processed to generate mechanical screening results.

[0007] S300. Based on the mechanical screening results, perform segmentation processing with a fixed time window size to obtain time window segmentation results; based on the time window segmentation results, perform time synchronization processing to generate synchronized screening results; based on the synchronized screening results, perform fault tag cross-comparison and comprehensive evaluation processing to generate comprehensive evaluation results.

[0008] S400. Based on the comprehensive evaluation results, alarm signals and maintenance suggestions are generated, and fault diagnosis results are generated.

[0009] Furthermore, the process of field aggregation and channel mapping includes:

[0010] The field aggregation process includes reading the sampling time range, acquisition batch mark, speed registration item and load condition registration item from the working condition group records one by one according to the working condition group; the channel correspondence process includes establishing a one-to-one correspondence between each sampled value and the acquisition channel name, sampling position, sampling time and working condition group number that generated the sampled value, and generating a working condition dataset.

[0011] Furthermore, the process of adjusting the bearing vibration frequency includes:

[0012] The bearing vibration frequency sorting process includes sorting the sampling time, registering the frequency range, marking abnormal frequencies, and merging continuous segments for all bearing vibration frequency records within the same operating condition group, forming a bearing vibration frequency sorting sequence.

[0013] Furthermore, the process of adjusting bearing speed and corresponding audio signal processing includes:

[0014] The bearing speed adjustment includes performing speed value reading, speed status registration, and fluctuation segment marking on the bearing speed at the same sampling time as the bearing vibration frequency adjustment sequence, thereby forming a bearing speed adjustment sequence.

[0015] The audio signal correspondence processing includes dividing the audio signals collected by the audio acquisition device in the same working condition group into corresponding audio segments according to the sampling time, and then establishing a correspondence between each corresponding audio segment and the bearing vibration frequency and bearing rotation speed at the same sampling time to form an audio correspondence record and generate mechanical screening results.

[0016] Furthermore, the process of segmenting data into fixed time window sizes includes:

[0017] The fixed time window segmentation process includes reading the demagnetization screening results according to the working condition group number, arranging the demagnetization candidate records, qualified records, and verification records in the order of sampling time, generating the first time window starting from the first valid sampling time, merging all records falling within the time window into the same segment record, generating subsequent time windows according to a fixed advancement rule, and repeating the merging process until all sampling times in the working condition group are segmented.

[0018] Furthermore, the time synchronization process includes:

[0019] The time synchronization process includes first reading the start time, end time, and operating condition group number of the time window segmentation results; then retrieving records that fall within the time window range from the winding screening results and the mechanical screening results; subsequently, performing a one-to-one correspondence process according to the sampling time, writing the demagnetization candidate records, winding candidate records, mechanical candidate records, and corresponding qualified records and review records within the same time window into the same synchronization record; when the winding screening results or the mechanical screening results are missing within a certain time window, a missing marker is written into the synchronization record.

[0020] Furthermore, the process of cross-referencing fault tags includes:

[0021] The cross-comparison of fault tags includes comparing demagnetization fault tags, winding fault tags, and mechanical fault tags within the same window. When two fault tags are consistent within the same time window and the other one is the same fault tag or a qualified tag, the time window is registered as a consistent record. When two fault tags are inconsistent within the same time window or one is a verification tag and the other two are inconsistent, the time window is registered as a conflict record. When only one fault tag is present within the same time window and the other two are qualified tags, the time window is registered as a single candidate record.

[0022] Furthermore, the comprehensive evaluation process includes:

[0023] The comprehensive evaluation process includes retaining all synchronized records within the time window for the consistent records and writing a comprehensive tag, retaining the original three-way states for the conflict records and writing a conflict tag, and retaining the candidate state for the single-way candidate record and writing a candidate tag.

[0024] Furthermore, the process of generating an alarm signal includes:

[0025] The alarm signal generation process includes reading the fault label and confidence score according to the working condition group number and time window number, then distinguishing between the source of comprehensive diagnostic results and the source of candidate records. For records in the source of comprehensive diagnostic results, when the fault label is clear and the confidence score reaches the preset threshold, the corresponding alarm signal is output.

[0026] Furthermore, the maintenance suggestion generation process includes:

[0027] The maintenance suggestion generation process includes, for records from the candidate record source, when a candidate fault label exists and the confidence score reaches a preset threshold, outputting a candidate alarm signal and simultaneously generating a retest-type maintenance suggestion; when a candidate fault label exists but the confidence score is lower than the preset threshold, not outputting a strong alarm signal, but generating a periodic maintenance suggestion or a retest-type maintenance suggestion; for records from conflict record sources, prioritizing the output of retest-type maintenance suggestions. The alarm signal generation process also includes continuous determination and repetition suppression. The continuous determination is used to check whether the same fault label appears consecutively within adjacent time windows. When it appears consecutively, the subsequent alarm signal is registered as a continuous alarm signal. The repetition suppression is used to prevent the repeated generation of the same type of alarm signal or repeated maintenance suggestion within the same time window, generating a fault diagnosis result.

[0028] The key innovations of this invention include:

[0029] (1) Based on the mechanical screening results, the time window segmentation results are obtained by first performing segmentation processing with a fixed time window size, and then time synchronization processing is performed based on the time window segmentation results to generate synchronous screening results, thereby incorporating the mechanical screening results into the continuous processing link under a unified time window.

[0030] (2) Based on the synchronous screening results, cross-comparison and comprehensive evaluation of fault labels are performed to generate comprehensive evaluation results, thereby using the record relationship under the same time window as the direct basis for subsequent processing.

[0031] (3) Based on the comprehensive evaluation results, alarm signals and maintenance suggestions are generated to generate fault diagnosis results, thereby establishing a correspondence between the comprehensive evaluation results and the fault diagnosis results.

[0032] The following are its main beneficial effects:

[0033] (1) In view of the problem that the connection between the time window segmentation results and the time synchronization processing is unclear in the existing scheme, the present invention makes the mechanical screening results form synchronous screening results under the same time window by setting the segmentation processing and time synchronization processing of a fixed time window size. The source of the preceding record of the fault diagnosis results is clearer and it is suitable for the continuous processing in the fault diagnosis scenario of permanent magnet motor.

[0034] (2) In view of the problem that the synchronous screening results and fault tag cross-comparison are not sufficiently correlated in the existing scheme, the present invention performs fault tag cross-comparison and comprehensive evaluation on the synchronous screening results, so that the comprehensive evaluation results directly inherit the previous synchronous screening results, the record relationship within the time window is preserved, and subsequent processing is carried out according to a unified link.

[0035] (3) In view of the problem that the source of the comprehensive evaluation results is unclear and the subsequent output relationship is unclear in the existing scheme, the present invention generates alarm signals and maintenance suggestions based on the comprehensive evaluation results, so that the fault diagnosis results and the comprehensive evaluation results are corresponding, which is convenient for calling and reviewing in equipment operation monitoring and subsequent maintenance operations. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating a deep learning-based fault diagnosis method for permanent magnet motors, as provided in an embodiment of this application. Detailed Implementation

[0037] Example 1: Refer to Figure 1 This is a flowchart illustrating a deep learning-based fault diagnosis method for permanent magnet motors provided in an embodiment of the present invention. The process may include at least steps S100-S400:

[0038] S100. Based on the rotor surface temperature value and the stator coil temperature value, field aggregation and channel correspondence processing are performed to obtain the working condition dataset;

[0039] S200. Based on the aforementioned working condition dataset, the bearing vibration frequency is sorted, the bearing speed is sorted, and the audio signal is processed accordingly to generate mechanical screening results.

[0040] S300. Based on the mechanical screening results, perform segmentation processing with a fixed time window size to obtain time window segmentation results; based on the time window segmentation results, perform time synchronization processing to generate synchronized screening results; based on the synchronized screening results, perform fault tag cross-comparison and comprehensive evaluation processing to generate comprehensive evaluation results.

[0041] S400. Based on the comprehensive evaluation results, alarm signals and maintenance suggestions are generated, and fault diagnosis results are generated.

[0042] Step S100 includes at least steps S110-S130:

[0043] S110: Acquire rotor surface temperature, stator coil temperature, stator coil resistance, stator coil voltage, three-phase stator current signal, zero-sequence voltage signal, bearing vibration frequency, bearing speed, vibration signal, audio signal, and air gap magnetic flux density, and perform acquisition time registration processing to obtain the acquisition time record.

[0044] Specifically, the rotor surface temperature value refers to the temperature acquisition value of the outer surface of the permanent magnet motor rotor at the sampling time corresponding to the operating state; the stator coil temperature value refers to the temperature acquisition value of each acquisition point of the designated sub-coil at the sampling time corresponding to the operating state; the stator coil resistance value refers to the resistance acquisition value of the designated sub-coil circuit at the sampling time; the stator coil voltage value refers to the voltage acquisition value of the designated sub-coil circuit at the sampling time; the three-phase stator current signal refers to the current sampling sequence of the corresponding channel of the three-phase winding; the zero-sequence voltage signal refers to the voltage sampling sequence corresponding to the zero-sequence channel; the bearing vibration frequency refers to the vibration frequency acquisition value of the bearing position vibration channel within the sampling period; the bearing speed refers to the speed acquisition value of the shaft at the sampling time; the vibration signal refers to the time-series sampling signal output by the vibration sensor; the audio signal refers to the operating sound signal output by the audio acquisition device; and the air gap magnetic flux density is the magnetic flux density acquisition value at the designated tooth position at the sampling time. The acquisition action is performed by the acquisition channel, which includes a thermal imager, a current acquisition channel, a voltage acquisition channel, an oscilloscope, a vibration sensor, an audio acquisition device, and a Hall sensor probe of a gaussmeter. A thermal imager is positioned on the outside of the permanent magnet motor housing, aligned with the rotor surface temperature acquisition area and the stator coil temperature acquisition area. After outputting the temperature image, the corresponding temperature acquisition value is read. The current acquisition channel is connected to the corresponding circuit of the three-phase stator current signal, and the voltage acquisition channel is connected to the corresponding circuit of the zero-sequence voltage signal and the stator coil voltage value. An oscilloscope is connected to the bearing speed sampling port, a vibration sensor is installed on the outer wall of the bearing housing, an audio acquisition device is installed in a fixed position on the outside of the machine base, and the Hall sensor probe of the gaussmeter is embedded in the stator teeth and faces the air gap.

[0045] Understandably, the data acquisition time registration process is executed immediately after a round of synchronous sampling is completed in the acquisition channel. The triggering conditions for synchronous sampling include equipment startup, arrival of a fixed time window, speed change exceeding a preset level, load condition switching, continuous change in temperature acquisition values, continuous change in vibration signals, and arrival of a manual retest instruction. The data acquisition time registration process does not simply write a timestamp; instead, it records the sampling date, sampling time, sampling order, sampling channel name, equipment operating status, and channel access status for each sampled value. Then, it attaches the same acquisition batch mark to all sampled values ​​in the same round of sampling. When a sampled value is missing, delayed, or shows a significant jump, the data acquisition time registration process synchronously records the abnormal channel name, abnormal time, and resampling status, and retains both the original record and the resampling record under the same acquisition batch mark, facilitating subsequent steps to group them according to the same acquisition batch.

[0046] Furthermore, in practical engineering scenarios, the permanent magnet motor can be placed at the drive end of a fan or pump. During continuous operation of the equipment, the controller triggers each acquisition channel according to a fixed time window, first reading the temperature acquisition value, then reading the current, voltage, vibration, audio, and air gap magnetic flux density, and finally writing them all into the acquisition time record. Thus, the acquisition time record is written into the record area as the output field name of this step, and also serves as the input field name for S120 for the acquisition time record to access. Simultaneously, the acquisition time record retains the preceding correspondence between each acquisition value and subsequent operating condition group records, operating condition datasets, demagnetization screening results, winding screening results, and mechanical screening results.

[0047] S120. Based on the collected time records, perform speed registration and load condition registration processing to generate condition group records.

[0048] Specifically, after the acquisition time record enters this step, it is first read according to the acquisition batch mark completion order, and then the bearing speed, three-phase stator current signal, stator coil voltage value, stator coil resistance value, rotor surface temperature value, and stator coil temperature value are extracted according to the sampling time. The speed registration refers to performing speed value reading, speed range matching, and speed status writing for the bearing speed in each acquisition time record to form a speed registration item that corresponds one-to-one with the sampling time.

[0049] The load condition registration refers to identifying the load condition of the three-phase stator current signal, stator coil voltage value, stator coil resistance value, and temperature acquisition value in the same acquisition time record, and writing the identification result into the load condition registration item corresponding to the same sampling time. The load condition can be recorded in four categories: no-load, light-load, rated load, and heavy-load, or it can be recorded using the existing load condition description on site, as long as it is consistent.

[0050] The speed registration and load condition registration are performed by the condition grouping unit, which can be located within the acquisition controller or within the host computer. The two are associated through acquisition batch markers and sampling times. In actual operation, the condition grouping unit first retrieves all sampling times in the acquisition time record, then completes the speed registration one by one, followed by the load condition registration one by one, and finally groups the sampling records of the same speed range and the same load condition into the same condition group.

[0051] When the bearing speed at the sampling time is stable, but the three-phase stator current signal and stator coil voltage value show signs of load switching, the operating condition grouping unit prioritizes retaining the original value of the sampling time and separately registers this sampling time as a transitional load operating condition record for separate processing during subsequent field aggregation. When the bearing speed, three-phase stator current signal, and temperature acquisition value at the sampling time are all within a stable range, the operating condition grouping unit includes this sampling time in the stable operating condition record. Furthermore, the operating condition grouping record not only saves the speed range and load conditions, but also the start and end relationships of the sampling time, the acquisition batch mark, the acquisition channel name, the abnormal channel status, and the re-sampling status.

[0052] Thus, during the subsequent field aggregation and channel correspondence processing in S130, all sampled values ​​and abnormal records within the same operating condition group can be directly extracted from the operating condition group records, without having to re-retrieve the original acquisition time records. Understandably, in actual engineering embodiments, if the permanent magnet motor driving the wind turbine switches from light load to rated load, the controller will generate different operating condition group records for the speed fluctuation segment, load increase segment, and stable operation segment respectively, avoiding mixing the operating condition switching stage and the stable operation stage into the same group. Therefore, the operating condition group records are written into the group record area as the output field name of this step, and are also used as the input field name for S130 for the operating condition group records to call. Simultaneously, the operating condition group records continue to connect to the operating condition dataset establishment process in S200.

[0053] S130. Based on the operating condition grouping records, perform field aggregation and channel correspondence processing to generate an operating condition dataset.

[0054] Specifically, after the operating condition grouping records enter this step, the field aggregation unit reads the sampling time range, acquisition batch mark, speed registration item, and load operating condition registration item from the operating condition grouping records group by group. Then, it retrieves the rotor surface temperature value, stator coil temperature value, stator coil resistance value, stator coil voltage value, three-phase stator current signal, zero-sequence voltage signal, bearing vibration frequency, bearing speed, vibration signal, audio signal, and air gap magnetic flux density at the corresponding sampling time. The field aggregation refers to writing the sampled values ​​within the same operating condition group, within the same sampling time period, and under the same acquisition batch mark into the same data record in a fixed field order. The channel correspondence processing refers to establishing a one-to-one correspondence between each sampled value and the acquisition channel name, sampling location, sampling time, and operating condition group number that generated the sampled value, forming a data structure that can be directly called.

[0055] The fields are arranged in a consistent order, prioritizing temperature, resistance, voltage, current, vibration, audio, and magnetic flux density. This facilitates the generation of demagnetization screening results by S210 for rotor surface temperature and air gap magnetic flux density, the generation of winding screening results by S220 for three-phase stator current signal, zero-sequence voltage signal, and stator coil acquisition values, and the generation of mechanical screening results by S230 for bearing vibration frequency, bearing speed, and audio signal.

[0056] Furthermore, field aggregation and channel mapping are not simply a matter of splicing. During processing, the field aggregation unit first checks whether there are missing items, duplicate items, or abnormal channel states in each field within the same working condition group; when there are missing items, the missing position is retained and a missing mark is written; when there are duplicate items, the next sampled value is retained according to the sampling time sequence, and the previous sampled value is moved to the supplementary sampling record; when there are abnormal channel states, the original sampled value is retained, and an abnormal channel state field is added.

[0057] After this processing, the subsequent generation processes of demagnetization screening results, winding screening results, and mechanical screening results can directly identify usable and abnormal sampled values, avoiding the mixing of sampled values ​​of unclear origin into the same operating condition group. Understandably, the operating condition dataset refers to a grouped data set after field aggregation and channel mapping by operating condition group. Its minimum field set includes rotor surface temperature value, stator coil temperature value, stator coil resistance value, stator coil voltage value, three-phase stator current signal, zero-sequence voltage signal, bearing vibration frequency, bearing speed, vibration signal, audio signal, air gap magnetic flux density, sampling time, acquisition batch mark, speed registration item, and load operating condition registration item. Among these, rotor surface temperature value, three-phase stator current signal, bearing vibration frequency, and air gap magnetic flux density constitute the core field set for generating subsequent screening results; abnormal channel status, supplementary sampling status, and acquisition channel name are preferred extended fields used for on-site verification and record tracking. In an operational engineering implementation, on-site personnel can select a certain working condition group on the host computer interface. The system will then retrieve all field records corresponding to that working condition group and write them into the working condition dataset storage area in a unified field order for direct access by S210, S220, and S230.

[0058] Therefore, the operating condition dataset is written into the dataset storage area as the output field name of this step, and is also called in subsequent steps as the input field name of S210, S220 and S230. This forms a continuous transmission relationship from the acquisition time record, operating condition group record to the operating condition dataset, and provides a unified data entry point for subsequent time window segmentation, time synchronization, fault label cross-comparison, comprehensive evaluation, diagnostic input dataset construction, deep learning model diagnosis, alarm signal generation and maintenance suggestion generation.

[0059] In summary, this step achieves the following technical benefits: First, it registers the acquisition time of the raw sampled values ​​from multiple acquisition channels, then registers the rotational speed and load conditions, and finally completes field aggregation and channel mapping according to the operating condition groups. This changes the processing path in existing technologies where multi-source data is directly preprocessed before entering the deep learning model. By first generating an operating condition dataset and then distributing it to the steps for generating demagnetization screening results, winding screening results, and mechanical screening results, the relationship between sampled values ​​and sampling channels within the same operating condition group can be fixed, facilitating subsequent time synchronization and fault label cross-comparison according to time windows.

[0060] Step S200 includes at least steps S210-S230:

[0061] S210. Obtain the operating condition dataset, process the rotor surface temperature value and the air gap magnetic flux density change, and obtain the demagnetization screening results.

[0062] Specifically, the operating condition dataset comes from the operating condition dataset generated by S130, and the operating condition dataset includes at least the rotor surface temperature value, air gap magnetic flux density, sampling time, acquisition batch marker, speed registration item, load operating condition registration item, and channel correspondence. The rotor surface temperature value sorting is performed by the demagnetizing screening unit, which can be set in the host computer, controller, or fault diagnosis device.

[0063] The rotor surface temperature value processing refers to reading all rotor surface temperature values ​​within the same operating condition group at the sampling time, and then performing grouping, outlier removal, missing value marking, and continuity checks according to the batch marking to form a rotor surface temperature sequence. Outlier removal prioritizes abnormal temperature values ​​caused by thermal imager field of view shift, glare interference, partial obstruction, and instantaneous jumps; missing value marking identifies records that existed at the sampling time but whose temperature values ​​were not recorded; and continuity checks check whether there are interruptions in recording between adjacent sampling times.

[0064] The process of organizing the air gap magnetic flux density changes involves reading the air gap magnetic flux density within the same operating condition group according to the sampling time corresponding to the rotor surface temperature sequence, aligning it according to the stator tooth position, sampling time, and batch marker, and then registering the changes in the air gap magnetic flux density records at adjacent sampling times to form an air gap magnetic flux density change sequence. Here, the air gap magnetic flux density change is not a single value, but a collection of magnetic flux density records changing with the sampling time within the same operating condition group. The records include the sampling time, stator tooth position, original air gap magnetic flux density value, change state, and channel correspondence. Further, the demagnetization screening unit performs correlation organization on the rotor surface temperature sequence and the air gap magnetic flux density change sequence. It first checks whether they belong to the same operating condition group, the same batch marker, and adjacent sampling times. Then, records with continuously increasing rotor surface temperature values ​​and continuously changing air gap magnetic flux density are included in the demagnetization candidate records; records with stable rotor surface temperature values ​​and stable air gap magnetic flux density are included in the qualified records; and records with missing or outlier value markers are included in the review records.

[0065] Understandably, when the permanent magnet motor is operating continuously under rated load, the demagnetization screening unit continuously reads the rotor surface temperature and air gap magnetic flux density from the operating condition dataset at a fixed sampling rate. When the load condition changes, the speed fluctuates significantly, or the thermal imager is recalibrated, the demagnetization screening unit first retains the original records and then separately organizes the records of the switching segment to avoid mixing short-term changes caused by the operating condition change into the stable operating condition segment. Through the above processing, the demagnetization screening result is written as the output field name into the screening result record area and can be called by the "Demagnetization Screening Result" of S310. At the same time, the demagnetization screening result retains the sampling time, operating condition group number, demagnetization candidate record, qualified record, and review record, which is convenient for subsequent segmented processing with a fixed time window size for direct reading.

[0066] S220. Based on the aforementioned operating condition dataset, perform three-phase stator current signal processing, zero-sequence voltage signal processing, and stator coil acquisition value correspondence processing to generate winding screening results.

[0067] Specifically, the operating condition dataset also comes from the operating condition dataset generated by S130. The three-phase stator current signal refers to the current sampling sequence formed by the three-phase windings during continuous operation, and the zero-sequence voltage signal refers to the voltage sampling sequence formed by the zero-sequence channel during continuous operation. The stator coil acquisition values ​​correspond to the stator coil acquisition values ​​in the processing, including the stator coil temperature value, stator coil resistance value, and stator coil voltage value. The winding screening unit reads the three-phase stator current signals within the same sampling period according to the operating condition group, first completing phase correspondence, sampling time alignment, missing segment marking, and abnormal segment marking, and then forming a three-phase stator current sorting sequence. The abnormal segments here mainly target current channel drift, sampling glitches, channel discontinuity, and load abrupt change segments.

[0068] Subsequently, the winding screening unit reads the zero-sequence voltage signal at the same sampling time as the three-phase stator current sorting sequence, completes zero-sequence channel correspondence, sampling time alignment, and continuity check, and forms a zero-sequence voltage sorting sequence. The stator coil acquisition value correspondence processing refers to establishing a correspondence between the stator coil temperature value, stator coil resistance value, and stator coil voltage value at the same sampling time and the three-phase stator current sorting sequence and zero-sequence voltage sorting sequence, generating a stator coil correspondence record. This correspondence includes at least sampling time correspondence, operating condition group correspondence, acquisition batch mark correspondence, and channel correspondence.

[0069] Furthermore, the winding screening unit jointly organizes the three-phase stator current processing sequence, the zero-sequence voltage processing sequence, and the corresponding records of the stator coils. Records showing current imbalance, abnormal zero-sequence voltage, increased stator coil temperature, or changed stator coil resistance at the same sampling moment are categorized into winding candidate records. Records with stable acquisition values ​​are categorized into qualified records. Records with missing channels, interrupted sampling, or conflicting acquisition values ​​are categorized into verification records. Here, a conflict in acquisition values ​​refers to an abnormality in the three-phase stator current signal and zero-sequence voltage signal at the same sampling moment, while the stator coil acquisition value shows no corresponding change, or the stator coil acquisition value has changed but the current and voltage channel records are incomplete.

[0070] Understandably, in a workable engineering embodiment, if the permanent magnet motor drives the water pump to run continuously, the winding screening unit can first read the three-phase stator current signal at each sampling cycle, then read the zero-sequence voltage signal, and subsequently call the stator coil temperature value, stator coil resistance value, and stator coil voltage value at the same sampling time to complete the corresponding processing. If a current path glitch is found at a certain sampling time, that sampling time is marked as a verification record and is not directly entered into the winding candidate record. Through the above processing, the winding screening result is written as an output field name into the screening result record area and is called by the "winding screening result" of S320; at the same time, the winding screening result retains the winding candidate record, qualified record, verification record, sampling time, and operating condition group number, forming a direct connection with subsequent time synchronization processing.

[0071] S230. Based on the aforementioned working condition dataset, the bearing vibration frequency is processed, the bearing speed is processed, and the audio signal is correspondingly processed to generate mechanical screening results.

[0072] Specifically, after the operating condition dataset enters this step, the mechanical screening unit extracts bearing vibration frequency, bearing speed, vibration signal, and audio signal according to the operating condition group. The bearing vibration frequency sorting refers to performing sampling time sorting, frequency range registration, abnormal frequency marking, and continuous segment merging on all bearing vibration frequency records within the same operating condition group to form a bearing vibration frequency sorting sequence. The abnormal frequency marking here mainly targets short-term abnormal records caused by bearing housing looseness, sampling jitter, external impact, and abnormal sensor contact. The bearing speed sorting refers to performing speed value reading, speed status registration, and fluctuation segment marking on the bearing speed at the same sampling time as the bearing vibration frequency sorting sequence to form a bearing speed sorting sequence. The audio signal correspondence processing refers to dividing the audio signal collected by the audio acquisition device within the same operating condition group into corresponding audio segments according to the sampling time, and then establishing a correspondence between each corresponding audio segment and the bearing vibration frequency and bearing speed at the same sampling time to form an audio correspondence record.

[0073] Furthermore, the mechanical screening unit performs joint processing on the bearing vibration frequency processing sequence, the bearing speed processing sequence, and the corresponding audio records. First, it checks whether the three belong to the same operating condition group, the same acquisition batch marker, and the same channel correspondence within the same sampling time. Then, records where bearing vibration frequency abnormalities, bearing speed fluctuations, and audio signal abnormalities occur simultaneously are classified into mechanical candidate records; records where all three are stable are classified into qualified records; and records where any one is missing or where sampling conflicts occur are classified into verification records. Here, abnormal audio signals refer to the presence of obvious abnormal noises, continuous friction sounds, intermittent impact sounds, or operating sound records that are significantly inconsistent with stable operating condition audio segments within the audio segment. Sampling conflicts refer to situations where the bearing vibration frequency has changed but the bearing speed and audio signal have not been registered synchronously, or where the corresponding audio record has a time offset.

[0074] Understandably, when the permanent magnet motor is installed at the main drive end of the wind turbine, the mechanical screening unit can automatically execute this step during continuous wind turbine operation. The oscilloscope outputs the bearing speed, the vibration sensor outputs the vibration signal and converts it into the bearing vibration frequency, and the audio acquisition device synchronously outputs the audio signal. Subsequently, the mechanical screening unit completes the corresponding processing within the same operating condition group. If there are external knocking or temporary maintenance sounds at the wind turbine site, the corresponding audio record will be retained, but an external interference flag will be written and included in the verification record, not directly mixed into the mechanical candidate record. Through the above processing, the mechanical screening result is written as the output field name into the screening result record area and can be called by the "mechanical screening result" of S320. At the same time, the mechanical screening result, the demagnetization screening result, and the winding screening result participate in time synchronization processing and fault tag cross-comparison in subsequent steps.

[0075] In summary, this step splits the operating condition dataset into three parallel processing chains: demagnetization screening results, winding screening results, and mechanical screening results. Instead of directly merging multi-source data and feeding it into the deep learning model, or screening in a fixed order, this step first completes the sorting, matching, and screening within the same operating condition group, and then sends the three types of screening results into subsequent time-synchronized processing. This allows records from different fault sources to be separated beforehand, facilitating subsequent cross-comparison of fault labels and comprehensive evaluation within the same time window.

[0076] Step S300 includes at least steps S310-S330:

[0077] S310. Obtain the demagnetization screening results, perform segmentation processing with a fixed time window size, and obtain the time window segmentation results.

[0078] Specifically, the demagnetization screening results come from S210, and the demagnetization screening results include at least the sampling time, operating condition group number, demagnetization candidate record, qualified record, and verification record. The fixed time window size segmentation processing is executed by the time window segmentation unit, which can be set in the host computer, controller, or fault diagnosis device, and maintains data connection with the demagnetization screening unit, winding screening unit, and mechanical screening unit.

[0079] The fixed time window size refers to the use of a uniform segment length to divide continuous sampling records within the same working condition group. The segment length remains unchanged in a single diagnostic task, with the segment start point aligned with the sampling time and the segment end point advancing sequentially according to the continuous sampling order.

[0080] During operation, the time window segmentation unit first reads the demagnetization screening results according to the working condition group number, then arranges the demagnetization candidate records, qualified records, and verification records in the order of sampling time. It then generates the first time window starting from the first valid sampling time and merges all records falling within that time window into the same segment record. Subsequently, the time window segmentation unit generates subsequent time windows according to a fixed advancement rule and repeats the merging process until all sampling times within the working condition group are segmented. This fixed advancement rule can be a continuous advancement method with adjacent time windows connected end-to-end, or a partially overlapping advancement method; the advancement method used remains consistent throughout the same operation.

[0081] Furthermore, the time window segmentation unit simultaneously checks the integrity of records within each time window during the segmentation process. When the same time window contains only demagnetization candidate records and qualified records, the time window is registered as a valid segment; when the same time window contains verification records or has missing sampling moments, the time window is registered as a verification segment; when adjacent sampling moments span different operating condition groups, the time window segmentation unit terminates the previous segment at the boundary of the operating condition group and restarts the segmentation from the next operating condition group, avoiding mixing records of different speeds and load conditions into the same time window.

[0082] Understandably, in practical engineering scenarios, if the permanent magnet motor operates continuously under rated load, the time window segmentation unit continuously reads the demagnetization screening results according to the sampling time, and retains the start time, end time, operating condition group number, and recorded status within each time window. If an operating condition switch occurs during operation, the last time window of the previous operating condition group ends at the switching point, and the time window for the next operating condition group is re-established from the first sampling time after the switching point. Through the above processing, the time window segmentation result is written as the output field name into the segmentation result record area and is called by the "time window segmentation result" of S320; at the same time, the time window segmentation result serves as the time reference for subsequent time synchronization processing, forming a direct connection with the winding screening result and mechanical screening result in the same main step.

[0083] S320. Based on the time window segmentation results, the winding screening results, and the mechanical screening results, perform time synchronization processing to generate synchronized screening results.

[0084] Specifically, the time window segmentation result comes from S310, the winding screening result comes from S220, and the mechanical screening result comes from S230. All three include the sampling time, the operating condition group number, and the corresponding candidate record, qualified record, or verification record.

[0085] The time synchronization process is performed by a time synchronization unit. The function of the time synchronization unit is to align the demagnetization screening results, winding screening results, and mechanical screening results to the same time reference within the same operating condition group and the same time window. Specifically, the time synchronization unit first reads the start time, end time, and operating condition group number of the time window segmentation results, and then retrieves the records in the winding screening results and the mechanical screening results that fall within the time window range.

[0086] Subsequently, the time synchronization unit performs a one-to-one correspondence processing according to the sampling time, writing the demagnetization candidate records, winding candidate records, mechanical candidate records, and their corresponding qualified records and verification records within the same time window into the same synchronization record. This time synchronization is not merely a matter of splicing time points; it simultaneously performs window correspondence, operating condition group correspondence, and record status correspondence. In other words, only records belonging to the same operating condition group and whose sampling times fall within the same time window are written into the same synchronization record; when the sampling times fall within the boundaries of adjacent time windows or belong to different operating condition groups, the time synchronization unit registers the corresponding records separately without merging them.

[0087] Furthermore, the time synchronization unit also performs exception handling during the synchronization process. When the winding screening result or the mechanical screening result is missing within a certain time window, the time synchronization unit retains the demagnetization record in the time window segmentation result and writes a missing mark in the synchronization record; when all three types of screening results exist, but one record is in the status of a verification record, the time synchronization unit retains the verification record in the synchronization record and registers the entire time window as a verification synchronization record; when the sampling times corresponding to the three types of screening results are slightly misaligned but still fall within the same time window, the time synchronization unit treats them as the same synchronization record according to the time window merging rule.

[0088] Understandably, in an operational engineering embodiment, if the demagnetization screening results show demagnetization candidate records, the winding screening results show qualified records, and the mechanical screening results show mechanical candidate records within a certain time window, the time synchronization unit writes these three types of records and their corresponding states into the same synchronization record. It does not perform deletion or overwriting in advance, but rather retains them for subsequent fault tag cross-comparison and comprehensive evaluation. Through the above processing, the synchronization screening results are written as output field names into the synchronization result record area and are available for use by S330's "Synchronization Screening Results." Simultaneously, the synchronization screening results connect the demagnetization screening results, winding screening results, and mechanical screening results into a unified record within the same time window, providing direct input for subsequent fault tag cross-comparison and comprehensive evaluation.

[0089] S330. Based on the synchronous screening results, perform cross-comparison and comprehensive evaluation of fault tags to generate comprehensive evaluation results.

[0090] Specifically, the synchronous screening results come from S320, and the synchronous screening results include at least the time window number, operating condition group number, demagnetization record status, winding record status, mechanical record status, and their respective candidate records, qualified records, and verification records.

[0091] The cross-comparison and comprehensive evaluation of fault tags are performed by the comprehensive evaluation unit. This unit can be set after the time synchronization unit or alongside the deep learning model diagnostic unit, but it must precede S410 in the data retrieval order. During operation, the comprehensive evaluation unit first registers fault tags for each synchronization record. Here, fault tags refer to the demagnetization fault tags, winding fault tags, and mechanical fault tags written for demagnetization screening results, winding screening results, and mechanical screening results, respectively. When a record is qualified, the corresponding qualified tag is written; when a record is a verification record, the corresponding verification tag is written.

[0092] Subsequently, the comprehensive evaluation unit performs cross-comparison within the same time window. This cross-comparison refers to comparing demagnetization fault tags, winding fault tags, and mechanical fault tags within the same window to determine whether they are consistent, conflicting, or whether a single candidate exists. Specifically, when at least two fault tags are consistent within the same time window, and the remaining tag is either the same fault tag or a qualified tag, the comprehensive evaluation unit registers this time window as a consistent record. When at least two fault tags are inconsistent within the same time window, or when one tag is a verification tag and the other two tags are inconsistent, the comprehensive evaluation unit registers this time window as a conflict record. When only one tag is a fault tag within the same time window, and the other two tags are qualified, the comprehensive evaluation unit registers this time window as a single candidate record.

[0093] Furthermore, after completing the cross-matching of fault labels, the comprehensive evaluation unit performs comprehensive evaluation processing. For consistent records, the comprehensive evaluation unit retains all synchronized records within the time window and writes a comprehensive label; for conflicting records, the comprehensive evaluation unit retains the original states of the three paths and writes a conflict label; for single-path candidate records, the comprehensive evaluation unit retains the candidate state of that path and writes a candidate label. This comprehensive evaluation processing does not simply select a particular filtering result, but rather completely saves the correspondence between the three filtering results within the same time window, allowing for direct use by the S410 when constructing the comprehensive dataset and candidate datasets.

[0094] Understandably, in actual engineering scenarios, if demagnetization fault tags and mechanical fault tags appear simultaneously within a certain time window, and the winding fault tag is a qualified tag, then the time window can be registered as a consistent record or a conflict record, depending on whether the demagnetization fault tag and the mechanical fault tag belong to the same fault path; if only the winding fault tag appears within a certain time window, and the other two paths are qualified tags, then the time window is registered as a single-path candidate record.

[0095] Through the above processing, the comprehensive evaluation result is written as the output field name into the evaluation result record area and can be called by the "Comprehensive Evaluation Result" of S410; at the same time, the comprehensive evaluation result is connected to the comprehensive dataset construction, candidate dataset construction, deep learning model diagnosis, fault label and confidence score generation, alarm signal generation and maintenance suggestion generation processing.

[0096] In summary, this step achieves the following technical benefits: It first segments the demagnetization screening results, winding screening results, and mechanical screening results into fixed time windows, then performs time synchronization and fault tag cross-comparison within the same time window. This changes the existing processing method of directly fusing multi-source data or sequentially screening in a fixed order. By retaining consistent records, conflicting records, and single-path candidate records within the same time window, the subsequent construction of the comprehensive dataset and candidate datasets has a clear source, and the boundaries of the input records received by the deep learning model for diagnosis are clearer.

[0097] In one embodiment, this step first obtains the demagnetization screening results from S210. These results include at least the sampling time, operating condition group number, demagnetization candidate records, qualified records, and review records. The time window segmentation unit reads the demagnetization screening results according to the operating condition group number, arranges all records within the same operating condition group in the order of sampling time, and then performs segmentation processing with a fixed time window size starting from the first valid sampling time of that operating condition group. The fixed time window size refers to using a uniform segment length to divide continuous sampling records within the same operating condition group. The segment length remains unchanged throughout a diagnostic task, the segment start point is aligned with the sampling time, and the segment end point advances sequentially according to the continuous sampling order. Specifically, during operation, the time window segmentation unit first reads the operating condition group number. Obtain the first valid sampling time within this operating condition group. And preset a fixed time window length. (For example, 60 seconds). For each sampling time within this operating condition group. Calculate the time window number to which it belongs, which is determined by formula ①:

[0098] Formula①

[0099]

[0100] in: The sampling time is directly read from the "Sampling Time" field of the demagnetization screening results; For operating condition group The first effective sampling time is extracted from the first demagnetization screening record of this working condition group; The time window length is fixed and preset by the diagnostic task. This is the floor function; The calculated time window number represents the sampling time. In the operating condition group Belongs to the first A time window.

[0101] Data source mapping: Extracting sampling time from "Demagnetization Screening Results" The result is determined by the "Operating Condition Group Number". The data was obtained from the first sampling time within this operating condition group. Preset parameters Provided by system configuration.

[0102] Simple numerical example: Assume a certain working condition group The first sampling time seconds, time window length seconds. If the sampling time seconds, then This means that the sampling time belongs to the third time window.

[0103] After obtaining the window number for each sampling time, the time window segmentation unit merges all demagnetization screening records (including demagnetization candidate records, qualified records, and review records) falling within the same time window into the same segment record, and simultaneously records the start time of that window. and end time ,in For window number, The operating condition group is numbered. During the segmentation process, the unit synchronously checks the integrity of the records within each time window: if the same time window contains only demagnetization candidate records and qualified records, the window is registered as a valid segment; if there are verification records or missing sampling times, it is registered as a verification segment; when adjacent sampling times cross different operating condition groups, the previous segment is terminated at the boundary of the operating condition group, and the segmentation restarts from the next operating condition group to avoid mixing records of different speeds and load conditions into the same time window. Through the above processing, the time window segmentation result is generated, which includes the start time of the time window, the end time of the time window, the operating condition group number, the record status within the window (valid segment or verification segment), and the corresponding window sequence number. This result is written as an output field name into the segmentation result record area and is called by the "Time Window Segmentation Result" of S320. At the same time, it serves as the time reference for subsequent time synchronization processing, forming a direct connection with the winding screening result and mechanical screening result in the same main step.

[0104] To facilitate understanding, a practical engineering example is provided: In a scenario where a permanent magnet motor-driven fan operates continuously under rated load, the demagnetization screening unit generates one record every 10 seconds (including sampling time, operating condition group number, and record type). Assume the operating condition group number is... The first sampling time seconds, time window length Within a 0-60 second window, if there are 6 records falling within the sampling timeframes of 10 seconds, 20 seconds, 30 seconds, 40 seconds, 50 seconds, and 60 seconds, then the 10-second and 20-second records are demagnetization candidate records, and the rest are qualified records. This window is then registered as a valid segment. Within a 60-120 second window, if there is a verification record (e.g., 80 seconds), this window is registered as a verification segment. When load switching causes a change in the operating condition group, the previous operating condition group ends at the switching point, and the window division restarts at the first sampling time after the load stabilizes for the next operating condition group. The final generated time window segmentation results will clearly specify the start and end times, operating condition group number, and window status for each window, for subsequent use.

[0105] Further, in S320, this step inherits the time window segmentation results generated in S310, and simultaneously acquires the winding screening results from S220 and the mechanical screening results from S230. All three include the sampling time, operating condition group number, and corresponding candidate records, qualified records, or verification records. The function of the time synchronization unit is to align the demagnetization screening results (here, the demagnetization screening results are directly referenced, not just the window segmentation results), winding screening results, and mechanical screening results to the same time reference within the same operating condition group and the same time window. Specifically, the time synchronization unit first reads the start time of the time window from the time window segmentation results. End time of the time window and operating condition group number Then, all records falling within the time window range are retrieved from both the winding screening results and the mechanical screening results. To aggregate multiple records that may exist within the window into a single state value, an aggregation function is defined. The rules are as follows: first, the candidate state (value 2) is selected; second, the verification state (value 1) is selected; third, the qualified state (value 0) is selected; and if there are no records, the missing state (value -1) is selected. This aggregation process is implemented by formula ②:

[0106] Formula②

[0107]

[0108]

[0109]

[0110] in: The time window number, Number the operating condition group; For time window The set of all sampling times within the time window is defined by the start and end times in the time window segmentation results; Indicates the sampling time in the demagnetization screening results Corresponding record types (candidate, review, qualified); and These represent the sampling times in the winding screening results and the mechanical screening results, respectively. The corresponding record type; For mapping functions, map record types to numeric values: Candidate Review ,qualified ;symbol Ensure that the maximum value is -1 when the set is empty, indicating a missing value; , , The aggregated numerical state has a range of values. These correspond to missing, qualified, reviewed, and candidate, respectively. Maximum value function; : Set construction symbol, representing the set of all elements that satisfy the condition following the vertical bar; Sampling time.

[0111] Data source mapping: Obtained from "Time Window Segment Results" Obtained from "Demagnetization Screening Results" Obtained from "Winding Screening Results" Obtained from "mechanical screening results" Preset mapping Defined internally by the system.

[0112] Simple numerical example: a window Within the demagnetization screening results, there are three records, with sampling times of 101 seconds (pass), 110 seconds (pass), and 115 seconds (review). The resulting value set is: The maximum value is 1, therefore (Verification). If there are no records in this window for winding screening results, the set is empty, and the maximum value is -1. (Missing). The mechanical screening results have one record (candidate) with a value of 2, therefore... (Candidate).

[0113] After obtaining the three convergence states, the time synchronization unit performs a one-to-one correspondence processing according to the sampling time, writing the demagnetization convergence state, winding convergence state, and mechanical convergence state within the same time window, along with the corresponding original records (optional), into the same synchronization record. This time synchronization is not merely time-based splicing, but simultaneously performs window correspondence, operating condition group correspondence, and record state correspondence: records are only written into the same synchronization record when they belong to the same operating condition group and their sampling times fall within the same time window; when sampling times fall within the boundaries of adjacent time windows or belong to different operating condition groups, they are registered separately and not merged. Furthermore, the unit performs exception handling during synchronization: when the winding screening result or mechanical screening result is missing within a certain time window, the demagnetization record in the time window segment results is retained, and a missing marker is written into the synchronization record; when all three types of screening results exist but one record state is "verification," the verification record is retained, and the entire time window is registered as a verification synchronization record; when the sampling times corresponding to the three types of screening results have slight misalignments but still fall within the same time window, they are considered the same synchronization record according to the time window merging rules. The above processing generates a synchronous filtering result, which includes at least the time window number, operating condition group number, and demagnetization record status. Winding record status Mechanical recording status This includes detailed data on the corresponding candidate records, qualified records, and review records (such as sampling time, specific fault type, etc.). This result is written as an output field name to the synchronization result record area and can be called by the "synchronization screening result" of S330, providing direct input for subsequent cross-comparison of fault labels and comprehensive evaluation.

[0114] Furthermore, in S330, this step receives the synchronous filtering result from S320, which includes the time window number. Operating condition group number Demagnetization Recording Status Winding record status Mechanical recording status And detailed records corresponding to each state (e.g., specific fault type). The comprehensive evaluation unit first performs fault tag registration for each synchronization record: according to If the value is 2 (candidate), then the specific fault type (such as "demagnetization fault") is extracted from the corresponding demagnetization candidate record as the demagnetization fault label. If the value is 1 or 0, it is registered as a review label or a pass label, respectively; similarly, a winding fault label is obtained. and mechanical failure labels Subsequently, the cells perform cross-comparison within the same time window, i.e., cross-comparison is performed on each cell. , , A comparison is performed among the three candidates to determine whether they are consistent, conflicting, or if a single candidate exists. This decision logic is represented by the comprehensive evaluation function in formula ③. Complete, output the overall evaluation type of this window. :

[0115] Formula③

[0116]

[0117] in, :window The comprehensive evaluation type is a string with values ​​including "consistent", "conflicting", and "single-path candidate". This type identifies the relationship between the three fault labels.

[0118] : Time window number, same as formula ②.

[0119] : Working condition group number, same as formula ②.

[0120] The comprehensive evaluation function has the following internal decision rules:

[0121] If at least two fault labels are the same and are specific fault types (unqualified, unverified), and the remaining one has the same fault type or a qualified label, then output "consistent";

[0122] If at least two fault labels are inconsistent and both are specific fault types (i.e., there are two different fault types), or one is a verification label and the status of the other two is inconsistent, then output "conflict";

[0123] If only one path is a specific fault type, and the other two paths are qualified labels, then output "single path candidate";

[0124] In other cases (such as cases involving review but not constituting the aforementioned conflict), "Pending Review" can be output or defined according to business needs.

[0125] The demagnetization fault label is obtained from the detailed fault type of the corresponding demagnetization candidate record in the synchronous screening results. If the demagnetization status in this window is qualified, then "qualified" is selected; if it is under review, then "review" is selected; if it is a candidate, then the fault description (such as "demagnetization fault") is extracted from the specific candidate record.

[0126] : Winding fault tag, similarly obtained from the winding records in the synchronous screening results.

[0127] Mechanical fault tags are similarly obtained from mechanical records in the synchronous screening results.

[0128] Data source mapping: Extract from "Synchronous Filter Results" , , And based on these states, retrieve the specific fault labels from the corresponding detailed records. , , .

[0129] Simple numerical example: a window The synchronous screening results show And the corresponding fault type is "demagnetization fault". And the corresponding fault type is "winding short circuit". (Pass). At this time "Demagnetization failure" "Winding short circuit" "Qualified" indicates that the two fault labels are inconsistent but both are specific fault types, thus meeting the conflict condition. "Conflict". If another window middle, "Demagnetization failure" "Demagnetization failure" If it is "qualified", then the consistency condition is met. "Consistent".

[0130] After completing the cross-comparison, the comprehensive evaluation unit performs comprehensive evaluation processing: for consistent records, all synchronized records within the time window are retained and a comprehensive tag is written (i.e., ("Consistent"); for conflict records, the original states of all three paths are preserved and a conflict label is written; for single-path candidate records, the candidate state of that path is preserved and a candidate label is written. This process does not simply select a particular filtering result, but rather completely saves the correspondence between the three filtering results within the same time window, forming a comprehensive evaluation result. The comprehensive evaluation result is written as an output field name to the evaluation result record area and is available for use by S410's "Comprehensive Evaluation Result," connecting to subsequent processes such as comprehensive dataset construction, candidate dataset construction, and deep learning model diagnosis.

[0131] In summary, the technical effects of this step are as follows: By first segmenting the three screening results into fixed time windows, and then completing time synchronization and cross-comparison of fault labels within the same time window, the processing method of direct fusion of multi-source data or fixed-order serial screening in the existing technology is changed. This makes the construction of subsequent comprehensive datasets and candidate datasets have clear sources, the boundaries of input records received by the deep learning model are clearer, and the interpretability and accuracy of diagnostic results are improved.

[0132] Step S400 includes at least steps S410-S430:

[0133] S410. Obtain the comprehensive evaluation results, perform comprehensive dataset construction and candidate dataset construction processing to obtain the diagnostic input dataset.

[0134] Specifically, the comprehensive evaluation result comes from S330, and the comprehensive evaluation result includes at least the time window number, operating condition group number, comprehensive label, conflict label, candidate label, demagnetization record status, winding record status, and mechanical record status. The comprehensive dataset construction and candidate dataset construction processes are executed by the diagnostic input construction unit, which can be set in a host computer, controller, or fault diagnosis device, and maintains a data connection with the comprehensive evaluation unit and the deep learning model. The comprehensive dataset refers to the data set formed after extracting consistent records and synchronization records that can be directly used for deep learning model diagnosis from the comprehensive evaluation result; the candidate dataset refers to the data set formed after extracting single-path candidate records, conflict records, and verification-related records from the comprehensive evaluation result. In actual operation, the diagnostic input construction unit first reads the comprehensive evaluation result according to the operating condition group number, and then sequentially retrieves the comprehensive label, conflict label, and candidate label corresponding to each time window according to the time window number.

[0135] For time windows registered as consistent records, the diagnostic input construction unit writes the demagnetization record status, winding record status, and mechanical record status within that time window, along with the corresponding sampling time, acquisition batch marker, speed registration item, and load condition registration item, into the comprehensive dataset. For time windows registered as single-channel candidate records, the diagnostic input construction unit writes the corresponding candidate status and the other two qualified statuses into the candidate dataset. For time windows registered as conflict records, the diagnostic input construction unit retains the three original statuses, conflict labels, and time window numbers, and writes them into the conflict record area of ​​the candidate dataset, without directly merging them into the comprehensive dataset.

[0136] Furthermore, the comprehensive dataset construction and candidate dataset construction processes also include field integrity checks, record order rearrangement, and input structure unification. Field integrity checks verify that the time window number, operating condition group number, demagnetization record status, winding record status, mechanical record status, and comprehensive label are complete. When a field is missing, the diagnostic input construction unit retains the corresponding time window in the candidate dataset and adds a missing field marker. Record order rearrangement arranges the time windows within the same operating condition group in chronological order, ensuring that the subsequent deep learning model reads data in the same order as the equipment operation. Input structure unification adjusts the field arrangement of each record in the comprehensive dataset and candidate dataset to a consistent format, allowing the deep learning model to use a unified entry point during reading.

[0137] Understandably, in real-world engineering scenarios, if multiple consecutive consistent records exist within a certain working condition group, these consistent records are directly written into the comprehensive dataset in time window order. If a single candidate record or conflicting record appears at the end of the same working condition group, these records are written into the candidate dataset and kept in their original time order within the diagnostic input dataset for subsequent processing by the deep learning model. Through the above processing, the diagnostic input dataset is written as the output field name into the diagnostic input storage area and is called by the "diagnostic input dataset" of S420. At the same time, the diagnostic input dataset maintains the internal distinction between the comprehensive dataset and the candidate dataset, enabling the subsequent fault label and confidence score generation processes to be executed separately according to the record source.

[0138] S420. Based on the diagnostic input dataset, perform deep learning model diagnostic processing to generate fault labels and confidence scores.

[0139] Specifically, the diagnostic input dataset comes from S410, and includes at least two parts: a comprehensive dataset and a candidate dataset. Both parts retain the time window number, operating condition group number, demagnetization record status, winding record status, and mechanical record status. The deep learning model diagnostic processing is executed by a deep learning model diagnostic unit, which includes a model reading part, a data input part, and a result output part.

[0140] The model reading section calls the deployed deep learning model, the data input section reads the diagnostic input dataset sequentially and feeds it into the deep learning model, and the result output section receives the diagnostic results from the deep learning model and writes them to the result record area. The deep learning model here can be a deep neural network model, a convolutional neural network model, or an autoencoder model, or a combination of the aforementioned models, but its established structure remains unchanged in this step. The key point is that the deep learning model reads the diagnostic input dataset after condition grouping, parallel filtering, time synchronization, and fault label cross-comparison. In actual operation, the deep learning model diagnostic unit first reads the comprehensive dataset and performs sequential input processing on each comprehensive record. Since the records in the comprehensive dataset come from consistent records, the deep learning model directly outputs the fault label and corresponding confidence score for this part of the records and writes the output to the comprehensive diagnostic result area.

[0141] Subsequently, the deep learning model diagnostic unit reads the candidate dataset and processes each candidate record sequentially. Since the records in the candidate dataset include single-path candidate records, conflict records, and records with missing fields, the deep learning model simultaneously reads the time window number, operating condition group number, and record status when processing these records, and outputs the candidate fault label and corresponding confidence score.

[0142] Furthermore, the deep learning model diagnostic unit performs input checks and result verification during processing. Input checks verify whether each record maintains a consistent field order. If a field order is found to be disordered, the record is returned to the diagnostic input storage area and marked for retesting, instead of being directly fed into the deep learning model. Result verification checks whether the fault labels output by the deep learning model are empty or whether confidence scores are missing. When the output is complete, it is written to the result record area; when the output is incomplete, the corresponding record is marked as a record to be retested. Understandably, in a workable engineering embodiment, during continuous equipment operation, the deep learning model diagnostic unit calls the diagnostic input dataset to perform diagnosis after the diagnostic input construction unit completes a batch of time window records, without waiting for the entire day's operation to end before processing. In this way, the comprehensive dataset can continuously output stable fault labels, and the candidate dataset can continuously output candidate fault labels and confidence scores, facilitating timely generation of alarm signals and maintenance suggestions in subsequent steps. Through the above processing, the fault label and confidence score are written as output field names into the diagnostic result storage area and can be called by S430's "fault label and confidence score"; at the same time, the fault label and confidence score continue to retain the correspondence with the time window number, operating condition group number and record source, thus forming a closed loop connection with the previous comprehensive evaluation results and subsequent fault diagnosis results.

[0143] S430. Based on the fault label and confidence score, perform alarm signal generation and maintenance suggestion generation processing to generate fault diagnosis results.

[0144] Specifically, the fault label and confidence score come from S420, and the fault label and confidence score include at least the time window number, operating condition group number, fault label, candidate fault label, comprehensive diagnostic result source, and candidate record source.

[0145] The alarm signal generation and maintenance suggestion generation process is executed by the result processing unit, which includes an alarm signal generation part, a maintenance suggestion generation part, and a result archiving part.

[0146] The alarm signal generation section outputs alarm signals based on fault labels and confidence scores, while the maintenance suggestion generation section outputs maintenance suggestions based on fault labels and confidence scores. The result archiving section integrates alarm signals, maintenance suggestions, and corresponding time window records into the fault diagnosis results. In practice, the result processing unit first reads the fault labels and confidence scores according to the operating condition group number and time window number, then distinguishes between the comprehensive diagnostic result source and the candidate record source. For records in the comprehensive diagnostic result source, when the fault label is clear and the confidence score reaches a preset threshold, the alarm signal generation section outputs the corresponding alarm signal and writes the fault label, alarm time, time window number, and operating condition group number into the result archiving section.

[0147] For records from the candidate record source, when a candidate fault label exists and the confidence score reaches a preset threshold, the result processing unit outputs a candidate alarm signal and simultaneously generates a retest-type maintenance suggestion; when a candidate fault label exists but the confidence score is lower than the preset threshold, the result processing unit does not output a strong alarm signal, but instead generates a periodic maintenance suggestion or a retest-type maintenance suggestion. For records from the conflict record source, the maintenance suggestion generation part prioritizes outputting retest-type maintenance suggestions, and retains the conflict label and the status of the three-way records in the fault diagnosis results, without directly deleting the original information.

[0148] Furthermore, the alarm signal generation and maintenance suggestion generation process also includes continuity determination and repetition suppression. Continuity determination is used to check whether the same fault tag appears consecutively within adjacent time windows; when it appears consecutively, the result processing unit registers the subsequent alarm signal as a continuous alarm signal. Repetition suppression is used to prevent the repeated generation of the same type of alarm signal or repeated maintenance suggestions within the same time window; when a corresponding record already exists within the same time window, the result processing unit only updates the record status and does not write a new record.

[0149] Understandably, in practical engineering scenarios, if the same fault label is output in three consecutive time windows within a certain operating condition group and the confidence score reaches a preset threshold, the result processing unit outputs an alarm signal in the first time window, outputs a continuous alarm status in the subsequent two time windows, and simultaneously provides maintenance suggestions for shutdown and repair. If a candidate record appears only in a single time window and has a low confidence score, the result processing unit generates maintenance suggestions for retesting or periodic maintenance, without directly triggering a strong alarm. Through the above processing, the fault diagnosis result is written as an output field name into the fault diagnosis result storage area and serves as the unified entry point for subsequent display, archiving, and manual review in this invention. The fault diagnosis result simultaneously retains the fault label, confidence score, alarm signal, maintenance suggestion, time window number, and operating condition group number, thus forming a complete correspondence with the preceding diagnostic input dataset and comprehensive evaluation results.

[0150] This step's technical effects can be summarized as follows: First, a comprehensive dataset and a candidate dataset are constructed. Then, a deep learning model processes the two types of records separately. Finally, alarm signals and maintenance suggestions are generated based on fault labels and confidence scores. This changes the existing technology's approach of directly diagnosing multi-source data and then uniformly outputting the results. By separating consistent records, candidate records, and conflicting records into the result processing flow, the source of the fault diagnosis results is clearer, and the alarm signals and maintenance suggestions maintain a direct correspondence with records in the preceding time window.

Claims

1. A fault diagnosis method for permanent magnet motors based on deep learning, characterized in that, include: S100. Based on the rotor surface temperature value and the stator coil temperature value, perform field aggregation and channel correspondence processing to obtain the operating condition dataset; the rotor surface temperature value refers to the temperature acquisition value of the outer surface of the permanent magnet motor rotor at the corresponding sampling time under the operating state; the stator coil temperature value refers to the temperature acquisition value of each acquisition point of the stator coil at the corresponding sampling time under the operating state. S200. Based on the aforementioned operating condition dataset, perform bearing vibration frequency processing, bearing speed processing, and audio signal correspondence processing to generate mechanical screening results; specifically including: S210. Obtain the operating condition dataset, process the rotor surface temperature value and the air gap magnetic flux density change, and obtain the demagnetization screening results. S220. Based on the operating condition dataset, perform three-phase stator current signal processing, zero-sequence voltage signal processing, and stator coil acquisition value correspondence processing to generate winding screening results. S230. Based on the aforementioned working condition dataset, the bearing vibration frequency is sorted, the bearing speed is sorted, and the audio signal is processed accordingly to generate mechanical screening results. S300. Based on the mechanical screening results, perform segmentation processing with a fixed time window size to obtain time window segmentation results; based on the time window segmentation results, perform time synchronization processing to generate synchronized screening results; based on the synchronized screening results, perform fault tag cross-comparison and comprehensive evaluation processing to generate comprehensive evaluation results. The fixed time window segmentation process includes reading the demagnetization screening results according to the working condition group number, arranging the demagnetization candidate records, qualified records and review records according to the sampling time sequence, generating the first time window from the first valid sampling time, merging all records falling within the time window into the same segment record, generating subsequent time windows according to a fixed advancement rule, and repeating the merging process until all sampling times in the working condition group are segmented. For each sampling time within this working condition group Calculate the time window number to which it belongs, which is determined by formula ①: Official ① in: Sampling time; For operating condition group The first effective sampling time; A fixed time window length; This is the floor function; The calculated time window number; The time synchronization process includes first reading the start time, end time, and operating condition group number of the time window segmentation results, then retrieving records that fall within the time window range from the winding screening results and the mechanical screening results, and then performing one-to-one correspondence processing according to the sampling time. The demagnetization candidate records, winding candidate records, mechanical candidate records, and corresponding qualified records and review records within the same time window are written into the same synchronization record. When the winding screening results or the mechanical screening results are missing in a certain time window, a missing mark is written into the synchronization record. The cross-comparison of fault tags includes comparing demagnetization fault tags, winding fault tags, and mechanical fault tags within the same window. When two fault tags are consistent within the same time window and the other one is the same fault tag or a qualified tag, the time window is registered as a consistent record. When two fault tags are inconsistent within the same time window or one is a verification tag and the other two are inconsistent, the time window is registered as a conflict record. When only one fault tag is within the same time window and the other two are qualified tags, the time window is registered as a single candidate record. The comprehensive evaluation process includes retaining all synchronized records within the time window for the consistent records and writing a comprehensive tag, retaining the original three-way states for the conflict records and writing a conflict tag, and retaining the candidate state for the single-way candidate record and writing a candidate tag. S400. Based on the comprehensive evaluation results, alarm signals and maintenance suggestions are generated, and fault diagnosis results are generated.

2. The method according to claim 1, characterized in that, The process of field aggregation and channel mapping includes: The field collection process includes reading the sampling time range, collection batch mark, speed registration item and load condition registration item from the working condition group records one group at a time according to the working condition group; The channel correspondence processing includes establishing a one-to-one correspondence between each sampled value and the acquisition channel name, sampling location, sampling time, and operating condition group number that generated the sampled value, thereby generating an operating condition dataset.

3. The method according to claim 1, characterized in that, The process of adjusting bearing vibration frequency includes: The bearing vibration frequency sorting process includes sorting the sampling time, registering the frequency range, marking abnormal frequencies, and merging continuous segments for all bearing vibration frequency records within the same operating condition group, forming a bearing vibration frequency sorting sequence.

4. The method according to claim 1, characterized in that, The process of adjusting bearing speed and corresponding audio signal processing includes: The bearing speed adjustment includes performing speed value reading, speed status registration, and fluctuation segment marking on the bearing speed at the same sampling time as the bearing vibration frequency adjustment sequence, thereby forming a bearing speed adjustment sequence. The audio signal correspondence processing includes dividing the audio signals collected by the audio acquisition device in the same working condition group into corresponding audio segments according to the sampling time, and then establishing a correspondence between each corresponding audio segment and the bearing vibration frequency and bearing rotation speed at the same sampling time to form an audio correspondence record and generate mechanical screening results.

5. The method according to claim 1, characterized in that, The process of generating an alarm signal includes: The fault label and confidence score are read according to the working condition group number and time window number. Then, the source of the comprehensive diagnosis result and the source of the candidate record are distinguished. For the record in the source of the comprehensive diagnosis result, when the fault label is clear and the confidence score reaches the preset threshold, the corresponding alarm signal is output.

6. The method according to claim 1, characterized in that, The maintenance suggestion generation process includes: The maintenance suggestion generation process includes, for records from the candidate record source, when a candidate fault label exists and the confidence score reaches a preset threshold, outputting a candidate alarm signal and simultaneously generating a retest-type maintenance suggestion. When a candidate fault label exists but the confidence score is lower than the preset threshold, a strong alarm signal is not output, but a regular maintenance suggestion or a retest-type maintenance suggestion is generated. For records from conflict record sources, a retest-type maintenance suggestion is output first. The alarm signal generation process also includes continuous determination and repetition suppression. The continuous determination is used to check whether the same fault label appears continuously within adjacent time windows. When it appears continuously, the subsequent alarm signal is registered as a continuous alarm signal. The repetition suppression is used to prevent the repeated generation of the same type of alarm signal or repeated maintenance suggestion within the same time window, generating a fault diagnosis result.

Images