Online monitoring and waveform recording method for multiple SEMIKRON DANFOSS SKIIP IPM modules

Abstract

The invention relates to the technical field of power electronic system monitoring, and discloses an online monitoring and waveform recording method for multiple SEMIKRON DANFOSS SKIIP IPM modules, and the method comprises the steps: collecting analog signals of each IPM module, and carrying out the multi-stage active filtering processing; performing synchronous sampling, signal amplification and anti-aliasing low-pass filtering processing on the analog signal to obtain digital data; dynamically adjusting a monitoring strategy of each IPM module according to the real-time operation state, performing data interaction with each IPM module based on a CANopen high-level communication protocol, and recording data displayed in a real-time waveform as a waveform file; and constructing a comprehensive health index to perform residual service life prediction, and realizing migration application on different batches or different models of IPM modules through a migration learning mechanism. According to the invention, the management level of the power electronic system is improved, and an effective technical means is provided for realizing predictive maintenance and intelligent operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of power electronic system monitoring technology, and more specifically, to an online monitoring and waveform recording method for multiple Semikrondanfos Skip IPM modules. Background Technology

[0002] With the rapid development of new energy power generation, high-voltage direct current transmission, high-performance industrial drives, and electric vehicle technologies, modern power electronic devices are constantly evolving towards higher power densities and more complex topologies. These devices typically employ multiple power modules connected in parallel, series, or hybrid topologies. As core components of power conversion systems, the operating status of these modules directly determines the performance, efficiency, and reliability of the entire system. Among them, intelligent power modules (IPMs), as integrated modules combining power switching devices, drive circuits, and protection circuits, are widely used in various power electronic systems. Currently, there are various models of intelligent power modules on the market. These modules typically possess parameter reading and writing capabilities and communication interfaces, providing a hardware foundation for system-level monitoring.

[0003] In actual operation, power modules endure electrical stress from high voltage, high current, and high-frequency switching operations, as well as the resulting intense thermal cycling, making them one of the most vulnerable components in the system. A failure in a single module can not only cause equipment downtime but also trigger a chain reaction, leading to catastrophic system failures, significant economic losses, and safety risks. Therefore, real-time, comprehensive, and synchronous monitoring and recording of the operating status of all critical power modules in the system has become an urgent requirement for ensuring the high reliability of power electronic equipment.

[0004] Currently, many traditional power module monitoring systems only monitor individual modules, making it difficult to meet the application needs of complex multi-module systems. In actual multi-module systems, the electrical and thermal stresses experienced by power modules in different locations vary significantly. Existing solutions are mostly limited to providing threshold alarms or monitoring single parameters such as temperature, lacking the ability to capture dynamic switching waveforms of voltage and current. This prevents engineers from accurately analyzing key issues such as overvoltage spikes, oscillations, and drive characteristics during the switching process. Furthermore, due to the lack of a black box function for multi-channel synchronous recording, fault analysis often relies on speculation, leading to recurring problems and severely reducing maintenance efficiency.

[0005] More importantly, the lifespan loss of power modules is the result of the combined effects of electrical and thermal stresses during operation. Electrical stress includes switching and conduction losses, while thermal stress includes junction temperature and temperature cycling. Traditional distributed independent monitoring schemes cannot obtain operational status data for each module with a unified timestamp, leading to a lack of comparability in the monitoring data and making it difficult to accurately assess the operational balance and health status of the entire system. Furthermore, the failure of one module in the system often triggers abnormal overvoltages or overcurrents in other modules. Without multi-channel waveform recording capabilities, it is impossible to fully capture the propagation paths and dynamic processes of voltage and current in the system when a fault occurs, making root cause analysis of system-level faults extremely difficult. Existing technologies also generally lack intelligent early warning functions, failing to predict the health degradation trend of modules in advance, and can only passively wait for faults to occur before taking action. This lag poses a significant threat to the safe operation of the system.

[0006] There are currently no effective solutions to the problems in the relevant technologies. Summary of the Invention

[0007] To address the problems in related technologies, this invention proposes an online monitoring and waveform recording method for multiple SEMIKRON DANFOSS SKIIP IPM modules, thereby overcoming the aforementioned technical problems in existing related technologies.

[0008] Therefore, the specific technical solution adopted by the present invention is as follows:

[0009] An online monitoring and waveform recording method for multiple SEMIKRON DANFOSS SKIIP IPM modules is disclosed. The method includes: acquiring voltage, current, and temperature signals from each IPM module as analog signals; performing multi-stage active filtering on the analog signals and interacting with each IPM module; synchronously sampling, amplifying, and performing anti-aliasing low-pass filtering on the multi-stage active filtering analog signals to obtain digital data, and transmitting the digital data using a cyclic redundancy check (CRC) mechanism; dynamically adjusting the monitoring strategy of each IPM module according to the real-time operating status; interacting with each IPM module based on the CANopen high-level communication protocol to receive the transmitted digital data and display the real-time waveform; recording the real-time waveform display data as a waveform file using a buffer writing strategy; extracting multi-dimensional features from the waveform file to construct a comprehensive health index, predicting the remaining service life, and implementing migration applications on different batches or models of IPM modules through a transfer learning mechanism.

[0010] Furthermore, the monitoring strategy for each IPM module is dynamically adjusted based on the real-time operating status, including: pre-defining a list of associated modules for each IPM module and setting a baseline sampling frequency for all IPM modules; this list of associated modules records other IPM modules that are electrically coupled or physically adjacent to the current IPM module; real-time monitoring of the characteristic parameters of each IPM module, including saturation voltage drop and temperature change rate; when the characteristic parameters of an IPM module exceed a preset abnormal threshold, the IPM module is immediately identified as an abnormal module; the sampling frequency of the abnormal module itself is increased to a high-frequency mode, and the sampling frequency of the associated modules of the abnormal module is increased to a medium-frequency mode according to the list of associated modules, while the sampling frequency of normal modules is reduced to a low-frequency mode to achieve intelligent scheduling of monitoring resources; when the characteristic parameters of the abnormal module return to normal levels and remain stable for a preset time, the sampling frequency of all IPM modules is restored to the baseline sampling frequency.

[0011] Furthermore, data interaction with each IPM module based on the CANopen high-level communication protocol includes: establishing an object dictionary mapping table for each IPM module instance, mapping the memory addresses of the fault status register, runtime counter, and temperature alarm threshold inside the IPM module to the CANopen object index and sub-index to achieve standardized access to the data inside the module; using synchronization messages as a global time base to synchronize the data acquisition time of all IPM modules; configuring real-time fault status data as event-driven process data object transmission; this real-time fault status data configuration is directly triggered by the setting transition of the fault status register inside the IPM module; configuring the runtime counter and temperature alarm threshold as periodic process data object transmission; and completing read and write operations on the non-volatile parameters of the IPM module based on the service data object segmented transmission protocol.

[0012] Furthermore, the read / write operations on the non-volatile parameters of the IPM module based on the Service Data Object Segmentation Transmission Protocol include: writing new operating parameters into the non-volatile memory inside the IPM module by sending a Service Data Object write request message; these operating parameters include temperature alarm threshold and fault protection delay; after the IPM module receives all data segments, it verifies the data segments; if the verification is correct, it stores the data segments in the non-volatile memory and replies with a Service Data Object write success response; if an error occurs during transmission, it replies with an abort message and initiates a retransmission; during the reading process, by sending a Service Data Object read request message, the IPM module transmits the parameters in the non-volatile memory in segments back; after receiving all the returned data segments, it verifies them to ensure data integrity.

[0013] Furthermore, data interaction with each IPM module based on the CANopen high-level communication protocol also includes rapid identification of the real-time fault status of the IPM module. Specifically, this includes: when an overcurrent or overtemperature fault occurs inside the IPM module, the state of the dedicated fault pin of the IPM module changes, the fault status register is set, and a pre-mapped process data object message is automatically sent; this process data object message does not require a host request; when an identifier from a faulty process data object is detected, the current task is immediately interrupted and the message is parsed; according to a predefined fault code mapping table, specific bytes of the error code in the process data object data content are matched with the mapping table to complete the determination of the fault level and fault type, and trigger the corresponding protection action; the fault level includes warning and critical fault, the fault type includes short circuit, drive fault and overheating, and the protection action includes popping up an alert window, recording the fault waveform and issuing a shutdown command.

[0014] Furthermore, the use of a cyclic redundancy check (CR) mechanism for digital data transmission includes: the data acquisition card, acting as the sender, calculates a first CR for the data packet before transmission and attaches the first CR to the data packet before sending; the PC monitoring software, acting as the receiver, receives the data packet, calculates a second CR for the data portion of the data packet, and determines whether the data transmission is correct by comparing the second CR with the first CR; if the second CR equals the first CR, the data is correct; if the second CR does not equal the first CR, the data packet is discarded and a retransmission is requested; the method for calculating the CR is as follows: the data to be checked is shifted left by 32 bits, a modulo operation is performed on a fixed generator polynomial, and the remainder is used as the check code.

[0015] The beneficial effects of this invention are as follows:

[0016] (1) This invention achieves synchronous acquisition of voltage, current and temperature data of multiple power modules through high-speed data acquisition card and dedicated sampling conditioning circuit. The dynamic time warping algorithm ensures that the signals from different modules can be strictly aligned to a unified time reference. High-frequency noise is filtered out by wavelet transform and data of different dimensions are normalized by Z-score method, thereby significantly improving the quality and reliability of monitoring data. The system also designs an intelligent collaborative monitoring mechanism, which can automatically adjust the sampling frequency according to the real-time operating status of each module. For modules that are detected to be abnormal, the sampling frequency is increased to 100kHz to capture transient processes. For modules that are electrically associated with abnormal modules, a medium frequency of 50kHz is used for monitoring. For modules that are operating normally, the sampling frequency is reduced to the basic frequency of 2kHz. This ensures fine monitoring of potential risk areas and avoids the waste of computing resources and storage space, solving the contradiction in traditional solutions where either the monitoring is not fine enough or the system is overburdened.

[0017] (2) This invention integrates predictive maintenance functions based on machine learning. By extracting the time-domain statistical features, frequency-domain spectral features, wavelet time-frequency features, and temperature-related thermal characteristic parameters of voltage and current waveforms, it uses principal component analysis to fuse these multidimensional features into a comprehensive health index. At the same time, it innovatively introduces a domain adaptation mechanism based on transfer learning. By defining the source domain and target domain, introducing a domain adaptation loss term, and establishing a cross-model model knowledge base, it significantly improves the generalization prediction ability of the model on different batches and different models of IPM modules, and can quickly adapt to new modules without starting training from scratch.

[0018] (3) This invention provides a complete multi-module intelligent monitoring solution. By using the CANopen communication protocol to access the status register and configuration parameters inside the power module, the object dictionary is used to map various monitoring data inside the module to standardized communication objects. Process data objects are used to realize efficient periodic data transmission and service data objects are used to complete parameter configuration and command issuance. The communication status of each module is detected in real time through the node monitoring mechanism, which meets the actual needs of high reliability applications such as new energy power generation and electric vehicle drive.

[0019] (4) The present invention adopts a multi-threaded architecture to separate and execute data acquisition, processing, display and storage tasks in parallel. It uses double buffering technology and OpenGL acceleration to realize smooth drawing of multi-channel waveforms. It uses CRC32 check to ensure the integrity of data transmission and storage. It also designs an automatic fault waveform recording function similar to that of an aircraft black box, which can automatically save the complete voltage and current waveforms before and after the fault when an anomaly is detected for post-fault analysis, thus providing a reliable technical means for the status monitoring and predictive maintenance of power electronic systems. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating an online monitoring and waveform recording method for multiple SEMIKRON DANFOSS SKIIP IPM modules according to an embodiment of the present invention.

[0022] Figure 2 This is a detailed implementation diagram of an online monitoring and waveform recording method for multiple SEMIKRON DANFOSS SKIIP IPM modules according to an embodiment of the present invention;

[0023] Figure 3 This is a schematic diagram of the process of data interaction between the IPM module and each IPM module based on the CANopen high-level communication protocol in an online monitoring and waveform recording method for multiple SEMIKRON DANFOSS SKIIP IPM modules according to an embodiment of the present invention.

[0024] Figure 4 This is a schematic diagram of the process of digital data transmission using a cyclic redundancy check mechanism in an online monitoring and waveform recording method for multiple SEMIKRON DANFOSS SKIIP IPM modules according to an embodiment of the present invention.

[0025] Figure 5 This is a schematic block diagram of an online monitoring and waveform recording system for multiple SEMIKRON DANFOSS SKIIP IPM modules according to an embodiment of the present invention. Detailed Implementation

[0026] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0027] According to an embodiment of the present invention, an online monitoring and waveform recording method for multiple SEMIKRON DANFOSS SKIIP IPM modules is provided.

[0028] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 and Figure 2 As shown, according to an embodiment of the present invention, an online monitoring and waveform recording method for multiple SEMIKRON DANFOSS SKIIP IPM modules is provided. This online monitoring and waveform recording method for multiple SEMIKRON DANFOSS SKIIP IPM modules includes:

[0029] The voltage, current, and temperature signals of each IPM module are collected as analog signals, multi-stage active filtering is performed on the analog signals, and data is exchanged with each IPM module.

[0030] The analog signal after multi-stage active filtering is synchronously sampled, amplified, and subjected to anti-aliasing low-pass filtering to obtain digital data, and a cyclic redundancy check mechanism is used for digital data transmission.

[0031] The monitoring strategy of each IPM module is dynamically adjusted according to the real-time operating status. Data interaction with each IPM module is carried out based on the CANopen high-level communication protocol to realize the reception of the transmitted digital data and the real-time waveform display. A buffer writing strategy is adopted to record the real-time waveform display data as a waveform file. Multi-dimensional features are extracted based on the waveform file to construct a comprehensive health index, predict the remaining service life, and realize the migration application on different batches or different models of IPM modules through a transfer learning mechanism.

[0032] In one embodiment, dynamically adjusting the monitoring strategy of each IPM module based on its real-time operating status includes: pre-defining a list of associated modules for each IPM module and setting a baseline sampling frequency for all IPM modules; the list of associated modules records other IPM modules that are electrically coupled or physically adjacent to the current IPM module; real-time monitoring of the characteristic parameters of each IPM module; the characteristic parameters include saturation voltage drop and temperature change rate; when the characteristic parameter of an IPM module exceeds a preset abnormal threshold, the IPM module is immediately identified as an abnormal module; the sampling frequency of the abnormal module itself is increased to a high-frequency mode, and the sampling frequency of the associated modules of the abnormal module is increased to a medium-frequency mode according to the list of associated modules, while the sampling frequency of normal modules is reduced to a low-frequency mode to achieve intelligent scheduling of monitoring resources; when the characteristic parameter of the abnormal module returns to a normal level and remains stable for a preset time, the sampling frequency of all IPM modules is restored to the baseline sampling frequency.

[0033] In one embodiment, data interaction with each IPM module based on the CANopen high-level communication protocol includes: establishing an object dictionary mapping table for each IPM module instance, mapping the memory addresses of the fault status register, runtime counter, and temperature alarm threshold inside the IPM module to the CANopen object index and sub-index to achieve standardized access to the data inside the module; using synchronization messages as a global time base to synchronize the data acquisition time of all IPM modules; configuring real-time fault status data as event-driven process data object transmission; this real-time fault status data configuration is directly triggered by the setting transition of the fault status register inside the IPM module; configuring the runtime counter and temperature alarm threshold as periodic process data object transmission; and completing read and write operations on the non-volatile parameters of the IPM module based on the service data object segmentation transmission protocol.

[0034] In one embodiment, the read / write operation of non-volatile parameters of the IPM module based on the Service Data Object Segmentation Transmission Protocol includes: writing new operating parameters into the non-volatile memory inside the IPM module by sending a Service Data Object write request message; the operating parameters include a temperature alarm threshold and a fault protection delay; after the IPM module receives all data segments, it verifies the data segments; if the verification is correct, it stores the data segments in the non-volatile memory and replies with a Service Data Object write success response; if an error occurs during transmission, it replies with an abort message and initiates a retransmission; during the reading process, by sending a Service Data Object read request message, the IPM module transmits the parameters in the non-volatile memory in segments back; after receiving all the returned data segments, it verifies them to ensure data integrity.

[0035] In one embodiment, such as Figure 3 As shown, data interaction with each IPM module based on the CANopen high-level communication protocol also includes rapid identification of the real-time fault status of the IPM module. Specifically, when an overcurrent or overtemperature fault occurs inside the IPM module, the state of the dedicated fault pin of the IPM module changes, the fault status register is set, and a pre-mapped process data object message is automatically sent. This process data object message does not require a host request. When an identifier from a faulty process data object is detected, the current task is immediately interrupted and the message is parsed. According to a predefined fault code mapping table, specific bytes of the error code in the process data object data content are matched with the mapping table to complete the determination of the fault level and fault type, and trigger the corresponding protection action. The fault level includes warning and critical fault, the fault type includes short circuit, drive fault and overheating, and the protection action includes popping up an alert window, recording the fault waveform and issuing a shutdown command.

[0036] In one embodiment, a multi-level fault filtering and dynamic priority parsing algorithm is used to determine the fault level and fault type. This includes: extracting fault codes from process data objects; comparing the fault codes with a validity check matrix to filter invalid error codes caused by communication interference; performing correlation analysis between the current fault and recent event logs; determining the fault level and the action to be triggered from a dynamic priority mapping table based on the fault codes and correlation analysis results; dynamically adjusting the dynamic priority mapping table according to the system operating mode, which includes start-stop, steady-state, and maintenance modes; when an IPM module reports a fault, querying the status of associated modules; if the associated module status is abnormal, determining it as a system-level fault and triggering system-level protection; if the associated module status is normal, triggering local isolation measures for the IPM module.

[0037] In one embodiment, such as Figure 4 As shown, digital data transmission using a cyclic redundancy check (CR) mechanism includes: the data acquisition card, acting as the sender, calculates a first CR for the data packet before transmission and attaches the first CR to the data packet before sending; the PC monitoring software, acting as the receiver, receives the data packet, calculates a second CR for the data portion of the packet, and determines whether the data transmission is correct by comparing the second CR with the first CR; if the second CR equals the first CR, the data is correct; if the second CR does not equal the first CR, the data packet is discarded and a retransmission is requested; the CR calculation method is as follows: the data to be checked is shifted left by 32 bits, a modulo operation is performed on a fixed generator polynomial, and the remainder is used as the check code; the expression for calculating the CR is:

[0038] C = (M << 32) mod P;

[0039] In the formula, C is the cyclic redundancy check code; M is the data message to be checked; and P is the fixed generator polynomial.

[0040] In one embodiment, receiving and displaying the transmitted digitized data in real time includes: converting the received raw data values ​​into screen pixel coordinates according to the range; appending newly arriving data points to the internal data container of the drawing component using incremental drawing technology; automatically scrolling the graphic view according to the new data range by setting a fixed time window to achieve a dynamic waveform effect; creating an off-screen image buffer in memory using a double buffering mechanism, where all drawing operations are completed in the buffer first, and the complete buffer content is exchanged to the screen for display only after a whole frame of graphics has been drawn to avoid flickering; assigning different drawing curve objects to voltage, current, and temperature and setting different colors and styles, with each curve sharing the same time axis and having an independent vertical axis and range to achieve synchronous comparison and display of multiple signals.

[0041] In one embodiment, extracting multidimensional features from waveform files to construct a comprehensive health indicator includes: using a dynamic time warping algorithm to align the time of multi-source signals in the high-frequency data of the waveform file; using a wavelet transform algorithm to denoise the signal and using a standardization method to normalize the data to obtain preprocessed data; extracting time-domain features, frequency-domain features, time-frequency-domain features, and thermal characteristic features from the preprocessed data as multidimensional features; the time-domain features include switching characteristic parameters and waveform statistics; the frequency-domain features include the amplitude of the dominant frequency component extracted by fast Fourier transform of the switching waveform; the time-frequency-domain features include the time-frequency energy distribution features extracted using wavelet basis; the thermal characteristic features include conduction loss, switching loss, junction temperature fluctuation, and mean junction temperature; using a principal component analysis algorithm to fuse and reduce the dimensions of the multidimensional features to construct a comprehensive health indicator, and labeling the preprocessed data with a remaining service life label.

[0042] In one embodiment, predicting remaining useful life includes: constructing a prediction model using a deep long short-term memory network; the prediction model includes an input layer, at least two long short-term memory hidden layers, a fully connected layer, and an output layer, used as input a historical feature sequence with continuous time steps, and outputting the predicted remaining useful life value of the IPM module; the prediction model calculates a loss function by comparing with the remaining useful life label, and uses a historical full life-cycle dataset of the IPM module accumulated from waveform files for supervised learning; the trained prediction model is embedded into PC-based monitoring software to perform feature extraction and model inference; the PC-based monitoring software displays the decay curve of comprehensive health indicators and the predicted remaining useful life value of the IPM module, and triggers an early warning based on a preset warning threshold; when the IPM module fails, the full life-cycle data of the IPM module accumulated from waveform files is used for incremental learning of the prediction model, and a sliding window time series update strategy is used to periodically train and optimize the prediction model.

[0043] In one embodiment, the transfer learning mechanism for applying the technology to IPM modules across different batches or models includes: using batches of IPM modules with accumulated complete remaining useful life (LHS) tag data as the source domain, and introducing newly introduced batches or models of IPM modules lacking tag information as the target domain; pre-training the prediction model on the source domain data, extracting common temporal features during the health degradation process of the IPM modules, and defining the output features of the second long short-term memory hidden layer in the pre-trained prediction model as deep shared features; inputting the temporal feature sequences from the source and target domains into the pre-trained prediction model, using the deep shared features to predict the remaining useful life and calculating the mean squared error loss, as well as the neighborhood distribution difference loss, and jointly optimizing the extended loss function. The system employs several methods to complete domain-adaptive training. Supervised fine-tuning is performed on the domain-adaptive training prediction model using early running data from the target domain. The network parameters of the input layer and the first long short-term memory hidden layer in the prediction model are frozen. The parameters of the second long short-term memory hidden layer, the fully connected layer, and the output layer in the prediction model are fine-tuned, and an early stopping strategy is adopted to obtain the target domain-adaptive prediction model. This early running data refers to the running data of the target domain IPM module at the beginning of its life cycle. When a new IPM module model is added, based on the electrical characteristics, rated parameters, and early signal characteristics of the new IPM module, the target domain-adaptive prediction model with the most similar feature distribution is selected from the cross-model model knowledge base as the transfer basis. The network structure and parameters of this target domain-adaptive prediction model are then loaded.

[0044] It should be noted that this invention discloses an online monitoring and waveform recording method for multiple SEMIKRON DANFOSS SKIIP IPM modules, aiming to solve the problems of existing technologies being unable to achieve synchronous high-speed acquisition of multiple modules, lagging fault early warning, and lack of intelligent health prediction. This invention uses the SEMIKRON DANFOSS SKIIP IPM module as a typical application object for technical solution verification. Since this type of module adopts a standardized communication protocol and interface design, the technical solution proposed in this invention is also applicable to other intelligent power modules with similar communication functions, which are collectively referred to as IPM modules for ease of description. This method achieves high-speed synchronous acquisition of voltage, current, and temperature signals from multiple IPM modules through the coordinated operation of sampling conditioning circuits, high-speed acquisition cards, and monitoring software; it adopts an intelligent collaborative monitoring mechanism that can dynamically adjust the sampling frequency according to the module's operating status; and it innovatively integrates a predictive maintenance framework based on machine learning, achieving accurate assessment of module health status and prediction of remaining lifespan through multi-dimensional feature extraction, health indicator construction, and LSTM prediction models. This invention significantly improves the reliability management level of power electronic systems and provides an effective technical means for achieving predictive maintenance and intelligent operation and maintenance.

[0045] like Figure 2As shown, this invention addresses the problems of lack of synchronization, insufficient waveform capture capability, and lack of intelligent early warning function in existing multi-IPM module monitoring systems. It provides a method for online monitoring and waveform recording of multiple SEMIKRON DANFOSS SKIIP IPM modules. The system includes a data acquisition card, a sampling conditioning circuit board, and a PC with monitoring software installed. The input terminals of the sampling conditioning circuit board are connected to the corresponding test points of multiple SEMIKRON DANFOSS SKIIP IPM modules. The input terminal of the data acquisition card is connected to the output terminal of the sampling conditioning circuit board. The monitoring software, deployed on the PC, can establish a communication connection with the data acquisition card via network cable or WiFi to receive data in real time and plot it as a dynamic waveform, while also supporting waveform recording and storage functions.

[0046] The sampling conditioning circuit board includes a sampling module, a filtering module, and a CAN communication module. The input of the circuit board is connected to the corresponding test points of multiple SEMIKRON DANFOSS SKIIP IPM modules through the sampling module.

[0047] This sampling module uses a high-precision operational amplifier circuit to acquire the voltage V of each SEMIKRON DANFOSS SKIIPIPM module in real time. ce Current I c and temperature T j The three types of analog signals are connected to the filtering module at the output. The acquired analog signals are filtered by the multi-stage active filters of the filtering module to effectively suppress noise and interference before being sent to the output of the sampling and conditioning circuit board.

[0048] Specifically, this method employs an intelligent collaborative monitoring mechanism that dynamically adjusts monitoring strategies based on real-time operating status. This includes: during the system configuration phase, pre-defining a list of associated modules for each module based on electrical topology and physical layout using monitoring software, and setting a baseline sampling frequency (e.g., 10kHz) for all modules for routine steady-state monitoring. The system monitors the characteristic parameters S of each module in real time, including Vce saturation voltage drop and temperature change rate. When a module's characteristic parameter S exceeds a preset abnormal threshold, the system will detect the abnormality. Upon detection, the module is immediately identified as an abnormal module, and a collaborative monitoring algorithm is activated: first, the sampling frequency of the abnormal module itself is increased to a high-frequency mode, such as 100kHz, while the sampling frequency of its associated modules is increased to a mid-frequency mode, such as 50kHz; for normal modules, the sampling frequency is appropriately reduced to a low-frequency mode, such as 2kHz. This dynamic resource allocation mechanism, through intelligent scheduling of monitoring resources, ensures both precise monitoring of risk areas and optimizes system resource utilization. When the characteristic parameter S of the abnormal module recovers to... After the sampling frequency of all modules is restored to the baseline strategy and remains stable for a preset period of time, such as 10 minutes, the system will automatically restore the sampling frequency of all modules to the baseline strategy.

[0049] The input of this CAN communication module is connected to the CAN interface of each SEMIKRON DANFOSS SKIIP IPM module via shielded twisted-pair cables, and the output is connected to a USB-CAN adapter on the PC via an electrically isolated CAN transceiver. The monitoring software, based on the CANopen high-level communication protocol, enables dedicated data interaction with each SEMIKRON DANFOSS SKIIP IPM module. For the specific functions of the SEMIKRON DANFOSS SKIIP IPM module, this invention pre-configures detailed device configuration files and establishes a dedicated CANopen object dictionary mapping table for each physical module instance in the monitoring software. This mapping table maps key memory addresses within the module, such as fault status registers, runtime counters, and temperature alarm thresholds, to standard CANopen object indices and sub-indices, enabling standardized access to data within the module.

[0050] By utilizing CANopen's SYNC synchronization message as a global time reference, strict synchronization of data acquisition times across all access modules is achieved, providing a foundation for subsequent year-on-year analysis of multi-module data. For data transmission, this invention employs an intelligent differentiation mechanism: high-priority data, such as real-time fault status, is configured for event-driven PDO transmission. Its triggering condition does not depend on data units or dimensions, but is directly triggered by changes in internal hardware status flags such as fault interrupt registers, ensuring microsecond-level latency. Regular monitoring data, such as runtime counts and temperature alarm thresholds, are configured for periodic PDO transmission, with the period flexibly set according to monitoring requirements, such as 100ms. This differentiation mechanism is based on the logical design of the module's internal firmware and is collaboratively managed by the monitoring software through pre-configured PDO mapping parameters. Furthermore, reliable read / write operations on the module's non-volatile parameters are achieved through the SDO segmented transmission protocol: the monitoring software reads factory-preset parameters by sending SDO request messages; through SDO write operations, new operating parameters, such as temperature alarm thresholds, are reliably written into the module's internal non-volatile memory in multiple data segments. The writing process includes verification and confirmation mechanisms to ensure the integrity and reliability of parameter configuration. The same applies to the reading process.

[0051] Specifically, the monitoring software's rapid identification of real-time fault states relies on a hardware and software collaborative interrupt-driven mechanism. The specific steps are as follows: When a fault such as overcurrent or overtemperature occurs within the module, the state of its dedicated fault pin changes, immediately setting a specific fault status register. This setting immediately triggers a pre-mapped PDO message, which is automatically sent without host request. The monitoring software's CAN receive thread is designed as a high-priority interrupt service routine. Once a CAN ID from a faulty PDO is detected, the current task is immediately interrupted, and the message is parsed. Based on a predefined fault code mapping table, the software matches specific bytes of the error code in the PDO data content with the mapping table. Within 1ms, it can determine the fault level (warning, critical fault, etc.) and the type (short circuit, drive fault, or overheating), triggering corresponding protection actions such as popping up an alert window, recording fault waveforms, and issuing shutdown commands.

[0052] In this invention, the real-time fault state is a state directly detected and locked by the internal hardware circuit of the module, specifically including: DC-Lmk voltage higher than DC-Lmk overvoltage threshold, PCB temperature higher than PCB temperature threshold, DCB temperature higher than DCB temperature threshold, load current higher than threshold, BOT input signal frequency too high, overlap time between BOT input signal and TOP input signal exceeding 3us, VCE voltage on BOT switch too high, TOP input signal frequency too high, overlap time between TOP input signal and BOT input signal exceeding 3s, VCE voltage on TOP switch too high, internal error of driver board, although the error is active, the switching signal is still active, and client VP voltage too low.

[0053] The non-volatile parameters of the module reliably read / write module via the SDO segmented transmission protocol refer to the following: the monitoring software reliably writes new operating parameters, such as temperature alarm thresholds and fault protection delays, into the module's internal non-volatile memory, such as Flash, by sending SDO write request messages. The writing process includes a verification and confirmation mechanism: after receiving all data segments and verifying their correctness, the module stores them in Flash and replies with an SDO write success response; if an error occurs during transmission, it replies with an abort message and initiates a retransmission, thereby ensuring the integrity and reliability of the parameter configuration. The reading process uses the same mechanism.

[0054] Specifically, unlike traditional technologies that use simple lookup tables, this invention proposes a multi-level fault filtering and dynamic priority parsing algorithm, including: First, extracting fault codes from emergency messages or PDOs, comparing them with a validity check matrix to filter out invalid error codes that may be caused by communication interference. Then, performing correlation analysis between the current fault and recent event logs; for example, if an over-temperature fault occurs on a module that just reported an over-current fault, its credibility and processing priority will be increased. Then, based on the fault code and the correlation analysis results, determining the final fault level and the action to be triggered from a dynamic priority mapping table. This mapping table can be dynamically adjusted according to the system's start-up, shutdown, steady-state, and maintenance modes; for example, in maintenance mode, the level of some faults will be downgraded to warning instead of direct shutdown. Finally, collaborative decision-making is performed. For multi-module systems, if a module reports a fault, the algorithm will instantly query the status of its associated modules, such as the current and temperature of adjacent bridge arm modules. If the associated modules also show abnormal symptoms, it is determined to be a system-level fault, triggering system-level protection; if it is only a single-point anomaly, local isolation measures are triggered.

[0055] The data acquisition card uses the NET5652 model, and this invention primarily utilizes its AI analog input function. The input terminal of the data acquisition card is connected to the output terminal of the sampling and conditioning circuit board via a DB37 interface, receiving multiple analog signals after conditioning and filtering. The acquisition card employs a 24-bit high-precision ADC to synchronously sample, amplify, and perform anti-aliasing low-pass filtering on the input multi-module voltage, current, and temperature signals. The processed data is then transmitted to the monitoring software on the PC via a gigabit Ethernet cable or 5GHz WiFi in the form of a custom protocol packet.

[0056] The data transmission verification mechanism of this invention is as follows: the data acquisition card, acting as the sender, calculates the data packet before transmission, and the verification code C is calculated according to the formula... The calculation involves M, where M is the data message to be verified, and P is a fixed generator polynomial 0x04C11DB7, using the IEEE 802.3 standard. Upon receiving the data, the monitoring software, acting as the receiver, also calculates C' from the data portion. It then compares C' with the received C value to determine if the data transmission is correct. If C' equals C, the data is error-free; if C' does not equal C, the data packet is discarded and a retransmission is requested. This invention employs this verification method to ensure the integrity of data transmission, providing a reliable data foundation for subsequent machine learning analysis.

[0057] This PC-based monitoring software features parameter setting, real-time waveform display, and data logging. It employs a multi-threaded architecture, utilizing a producer-consumer model and a circular buffer for efficient data stream processing. During operation, a dedicated network receiving thread receives data and stores it in the circular buffer, while the UI display thread reads data from the buffer for real-time display. Waveform display utilizes incremental plotting and double buffering technology to achieve smooth, flicker-free, real-time dynamic display of multi-channel signals for voltage, current, and temperature. Each signal can have an independent range. Users can configure parameters such as the analog signal type, maximum sampling frequency of 1MHz, number of sampling points (1024-8192), and continuous or triggered sampling modes for each input channel through the software interface. Users can also specify the signal name, unit, and input range for each channel.

[0058] It should be added that the producer-consumer model in this invention is the core software architecture for achieving efficient, real-time data stream processing, including producer threads, buffers, consumer threads, and multi-threaded extensions. The producer thread, which is the network receiving thread of the data acquisition card, acts as the sole producer, continuously reading multi-channel voltage, current, and temperature data packets from the data acquisition card after synchronous sampling and digitization. The buffer uses a circular buffer as a shared memory area. The producer thread writes the received valid data packets sequentially into the circular buffer. The size of the circular buffer can accommodate several seconds of high-frequency data to smooth out instantaneous fluctuations in the data stream and prevent data loss due to temporary delays in consumer processing. The consumer thread is the user interface display and data recording thread. This thread reads the latest data from the circular buffer at a fixed rhythm, such as synchronized with the screen refresh rate. Read operations do not block producer writes, thus decoupling acquisition and display. Multi-threaded extension means that in the complex application of this invention, there may be more than one consumer. For example, there may be independent fault waveform trigger recording threads and health feature extraction threads: when a fault PDO message is detected, the fault waveform trigger recording thread, as a high-priority consumer, extracts data from the buffer for a specific time window before and after the fault and saves it to a file; the health feature extraction thread periodically reads data from the buffer or the stored waveform file, performs feature calculations, and provides them for the prediction model; these threads all act as consumers, safely reading data from the same or different dedicated buffers.

[0059] Specifically, in the producer thread, each arriving data packet undergoes an immediate CRC check. Packets that fail the check are discarded and do not enter the circular buffer, ensuring the integrity of the data in the buffer and providing a reliable data source for all subsequent consumers. When the system dynamically adjusts its monitoring strategy based on real-time operating status, such as increasing the sampling frequency of an abnormal module to 100kHz, the data throughput of the producer thread will increase instantaneously. A well-designed circular buffer can absorb this sudden surge in traffic, preventing data overflow, while ensuring that the consumer thread can still run smoothly without being stuck due to data congestion. The consumer thread uses a double-buffering mechanism for drawing. It consumes data from the circular buffer, converts it into coordinate points, completes the drawing of one frame of graphics in the off-screen buffer in memory, and then quickly swaps it to the screen for display. This process runs parallel to the producer's continuous filling of the data ring, jointly achieving a balance between high-speed acquisition and smooth display.

[0060] In this invention, the real-time waveform display drawing mechanism includes: first, receiving the raw data values, such as int16... t The data is converted to screen pixel coordinates based on the measurement range. Incremental plotting technology is employed, appending only newly arriving data points to the internal data container of the plotting component, and setting a fixed time window (e.g., displaying data from the last 5 seconds). The graphical view automatically scrolls according to the new data range, achieving a dynamic waveform effect. A double-buffering mechanism is also used, creating an off-screen image buffer in memory. All plotting operations are first completed in the buffer, and only after a complete frame of graphics is drawn is the entire buffer content swapped to the screen for display, effectively avoiding flickering and improving plotting efficiency. The software assigns different plotting curve objects to voltage, current, and temperature, setting different colors and styles. Each curve shares the same time axis but can have an independent Y-axis and range, thus enabling synchronous comparison and display of multiple signals.

[0061] The waveform recording file format is CSV, which can be directly opened and analyzed using tools such as Excel and MATLAB. The recording steps are as follows: When the user clicks the record button, the software creates a new file on the hard drive and writes a header to the file. The software appends the parsed and verified valid data packets obtained from the network receiving thread to the end of the file in a predefined binary format. A buffered write strategy is used: data is first written to a memory buffer, and then written to the hard drive all at once when the buffer is full or recording stops. This reduces the number of disk I / O operations, greatly improves recording performance, and avoids data loss. When the user clicks the stop button, the software closes the file. Simultaneously, users can choose to save the current software settings parameters, such as range and sampling rate, as a separate configuration file and associate it with the data file.

[0062] Compared to traditional threshold-based alarms or simple trend analysis, this invention integrates a machine learning-based predictive maintenance framework. This framework preprocesses the collected high-frequency data through DTW alignment, wavelet denoising, and Z-score normalization, extracting multi-dimensional features including time domain, frequency domain, time-frequency domain, and thermal characteristics. It then uses PCA fusion to construct a comprehensive health index (HI) and employs an LSTM model for RUL prediction. Specifically, this includes: firstly, continuously acquiring voltage, current, and temperature data from the SEMIKRON DANFOSS SKIIP IPM module using a multi-channel high-speed acquisition system with a sampling rate ≥100kHz; secondly, using the Dynamic Time Warping (DTW) algorithm to achieve strict time alignment of multi-source signals; thirdly, applying the wavelet transform algorithm based on the Daubechies wavelet basis for signal denoising; and finally, using Z-score normalization to normalize the data, expressed as follows: ,in The mean, The standard deviation is used to comprehensively extract time-domain features from the preprocessed data, including switching characteristic parameters such as turn-on delay time, turn-off delay time, rise time, fall time, and conduction voltage drop, as well as waveform statistics including mean, variance, kurtosis, and skewness. Frequency-domain features are also extracted, specifically by performing an FFT transform on the switching waveform to extract the amplitudes of the top five dominant frequency components. Time-frequency domain features are further extracted using the Morlet wavelet basis to extract time-frequency energy distribution characteristics. Finally, thermal characteristics are extracted, including the calculation of conduction loss. Switching losses The junction temperature fluctuation was estimated using the Foster thermal network model. and mean In the formula, Conduction loss refers to the steady-state power loss caused by the on-state resistance of the insulated-gate bipolar transistor or freewheeling diode in the SEMIKRONDANFOSS SKIIP IPM module when it is fully turned on. The voltage drop between the collector C and emitter E of the internal IGBT of the SEMIKRON DANFOSS SKIIP IPM module when it is in saturation conduction state; This is the collector current, which is the real-time current flowing through the collector of the IGBT in the SEMIKRONDANFOSS SKIIP IPM module. Switching losses refer to the power losses caused by the overlap of voltage and current during the turn-on and turn-off transients of the IGBT in the IPM module. This refers to the instantaneous voltage between the collector and emitter of the IGBT during the switching transient process. Finally, principal component analysis (PCA) is used to fuse and reduce the dimensionality of the above multidimensional features to construct a comprehensive health index (HI).

[0063] The scores of the first principal components are normalized to the [0,1] interval using a linear mapping formula to quantify the module decay process, and the remaining useful life is labeled on the data. The linear mapping formula is as follows:

[0064] ;

[0065] In the formula, HI is the health index, a dimensionless scalar between 0 and 1, used to quantify the health status of the IPM module; PC1 is the first principal component score, which is the projection value on the first principal component obtained by fusing and reducing the dimensionality of the extracted multi-dimensional features including time domain, frequency domain, time-frequency domain and thermal characteristics through the principal component analysis algorithm. It is used to represent the comprehensive time series that best reflects the trend of module performance degradation. This first principal component is the component with the largest variance contribution. and These are the minimum and maximum values ​​of all PC1 score sequences, respectively, within the historical full-lifecycle dataset used for model training or within a sufficiently long reference time window;

[0066] The expression for the remaining useful life label is:

[0067] ;

[0068] In the formula, RUL represents the remaining service life, which is the remaining operating time from the current moment until the expected failure time of the module; T end The failure point is the point in time when the IPM module is determined to have failed or reached the end of its life based on historical data or accelerated life testing; T current The current time refers to the time point at which the data record was collected.

[0069] A prediction model is constructed using a deep long short-term memory (LSTM) network. The network structure includes an input layer with 64 neurons, two LSTM hidden layers with 128 neurons each, a fully connected layer with 64 neurons, and an output layer with 1 neuron. It takes a historical feature sequence of 100 consecutive time steps as input and outputs the predicted RUL value. Supervised learning is performed using a historical full lifecycle dataset.

[0070] Training employs the mean squared error loss function and the Adam optimizer. The trained model is then converted to a lightweight format using TensorFlowLite and embedded into the monitoring system for daily automatic feature extraction and model inference. The expression for the mean squared error loss function is:

[0071] ;

[0072] In the formula, MSE is the mean squared error, a loss function value used to measure the difference between the model's predicted value and the true value; the smaller the value, the higher the prediction accuracy of the model; n is the number of samples, referring to the total number of data points in a set participating in the current loss calculation; ypred is the model's predicted value, that is, the predicted value of the remaining lifetime of the IPM module output by the prediction model based on the input historical feature sequence; ytrue is the true label value, that is, the true value of the remaining lifetime of the IPM module corresponding to the input feature sequence, which is determined in advance based on historical or test data.

[0073] The specific training process of the prediction model in this invention is as follows: First, the multidimensional features extracted from the waveform file and the corresponding remaining useful life labels constitute the training dataset. The historical feature sequence of 100 consecutive time steps is used as the model input, and the remaining useful life value corresponding to that moment is used as the expected output. The model continuously adjusts the network parameters through the backpropagation algorithm, gradually reducing the mean square error between the predicted value and the actual remaining useful life label. After training, the model can learn the temporal evolution pattern of the IPM module gradually degrading from a healthy state to failure. When a new feature sequence is input, a single neuron in the model's output layer will provide the corresponding remaining useful life prediction value. This prediction value represents the estimated time the module can continue to operate normally in the current healthy state, in hours. The monitoring software stores the trained model file locally and automatically loads the model parameters each time it starts. After receiving new waveform data and completing feature extraction, it inputs the feature sequence of the most recent 100 time steps into the model for inference calculation, obtains the remaining useful life prediction value at the current moment, and updates and displays it on the interface in real time.

[0074] The software interface visually displays the HI decay curve and predicted RUL value, and allows setting early warning thresholds. Hours or The system triggers early warnings in a timely manner; it also has a feedback mechanism designed so that when a module eventually fails, its entire lifecycle data is automatically used for incremental model learning. The sliding window time series update strategy is used to periodically retrain and optimize the model, so that the model can continuously adapt to new operating environments and device batches, thereby continuously improving prediction accuracy.

[0075] To enhance the model's generalization prediction capability across different batches and models of IPM modules, this invention introduces a domain adaptation mechanism based on transfer learning. Specifically, by defining source and target domains, a domain adaptation loss, such as MMD, is introduced onto the pre-trained LSTM model to jointly optimize prediction error and distribution differences. Fine-tuning is then performed using a small amount of early data from the target domain, establishing a cross-model model knowledge base to achieve accurate lifetime prediction and rapid capability transfer for different IPM module models. After online deployment, this framework can automatically perform health prediction and early warning, and continuously optimize the model through incremental learning after module failure. Specifically, it includes:

[0076] First, define the source and target domains. The IPM module batches that have accumulated complete lifetime tag data are used as the source domains; newly introduced different batches or models of modules that lack tag information are used as the target domains.

[0077] It should be noted that the transfer learning mechanism in this invention solves the problem of insufficient historical failure data for new models or batches of IPM modules, making it impossible to directly train a prediction model. Traditional methods require waiting for the new module to run through at least several complete lifecycles to accumulate sufficient training data, a process that can take several years. Transfer learning, however, allows the system to utilize the large amount of historical data accumulated from existing older modules as the source domain, transferring the general health degradation patterns contained therein to new modules. Although the rated parameters and specific electrical characteristics of different module models differ, they all exhibit common degradation trends during aging, such as extended switching time, increased junction temperature, and increased losses. These common patterns can be learned through a pre-trained model and retained in the lower-level parameters of the network. Then, by fine-tuning the higher-level parameters of the network using a small amount of early operating data from the new module, a prediction model suitable for the new module can be quickly established.

[0078] Secondly, an LSTM model is pre-trained on source domain data to extract common temporal features during the health degradation process of IPM modules. The output features of the second LSTM layer of the pre-trained model are defined as deep shared features, which carry common distribution information in the source and target domains. These common temporal features refer to the common, time-series-related pattern features exhibited by different batches or models of IPM modules during performance degradation. These features are not specific data points, but rather representational patterns abstracted from massive amounts of data by the deep learning model. They mainly include: common degradation trends, such as the gradual increase in on / off times (tr and tf) with aging; common fluctuation characteristics, such as the changes in the amplitude and frequency of junction temperature fluctuation ΔTj when the health status declines; and common correlation evolution, such as the changes in the coupling relationship between multiple feature parameters, for example, the pattern of how the correlation strength between the increase in junction temperature Tj and the increase in on-state voltage drop Vce changes over time under the same load.

[0079] Then, a domain-adaptive loss term is introduced during the model's domain-adaptive training phase. The specific logical flow is as follows: The temporal feature sequences of the source and target domains are input into the pre-trained LSTM model; the output of the second LSTM layer of this model is defined as a deep shared feature; this feature is used for two purposes simultaneously: first, to continue forward propagation to predict the RUL and calculate the mean squared error loss (MSE); second, to input the deep shared features of the source and target domains into a domain difference calculation module, using maximum mean difference (MMD) or adversarial training to calculate the domain distribution difference loss (Domain Loss). The following extended loss function is jointly optimized: ,in These are hyperparameters used to balance the contributions of prediction accuracy and domain fit. Domain-invariant feature learning is achieved through gradient inversion layers or constrained optimization, forcing the feature extraction network to ignore domain-specific information and focus on health degradation representations shared across domains.

[0080] Subsequently, supervised fine-tuning of the aforementioned domain-adaptive trained model was performed using a small amount of early running data from the target domain. The specific steps are as follows: ① Data preparation and preprocessing: collecting running data from the initial period of the new module's lifecycle in the target domain, such as the first 5%-10% of the lifecycle data, and estimating or assigning preliminary remaining lifecycle RUL labels based on the initial trend of performance degradation or the reliability data provided by the manufacturer, forming a target domain fine-tuning dataset; ② Combining freezing and fine-tuning: freezing the network parameters of lower layers in the model, such as the first LSTM layer and earlier layers; maintaining its learned common feature extraction ability unchanged, and only fine-tuning the parameters of higher layers in the model, such as the second LSTM layer and fully connected layers, so that the model can adapt to the unique degradation dynamics of the new module; ③ Fewer-shot learning and early stopping mechanism: using a smaller learning rate to train on the target domain data to avoid destroying existing feature representations; at the same time, adopting an early stopping strategy to terminate training when the verification loss no longer decreases, preventing overfitting and ensuring the model's generalization ability.

[0081] The adapted prediction models for different module models refer to the mature prediction models that are suitable for specific module models after the aforementioned domain adaptation and target domain fine-tuning. These models are stored in a cross-model knowledge base, recording their corresponding module models, key feature distributions, and performance indicators.

[0082] Finally, a cross-model knowledge base is established. When a new module model is added, the system automatically selects the most similar adapted model from the knowledge base as the transfer basis based on the new module's electrical characteristics, rated parameters, and early signal characteristics. Its network structure and parameters are then loaded, and the aforementioned fine-tuning steps are performed again. This method significantly reduces the cold-start cost of training a model from scratch for a new module and enables rapid and efficient transfer of predictive capabilities between different models.

[0083] It's worth noting that this cross-model model knowledge base stores the prediction model files corresponding to each adapted IPM module model. This includes the model's network structure definition and trained weight parameters, as well as key electrical characteristic parameters of the corresponding module, such as rated voltage range, rated current range, and switching frequency range, and the model's prediction accuracy on the validation set. When a new IPM module model is connected to the system, the monitoring software first reads the module's nameplate parameters and early operating data to calculate the statistical characteristics of its feature distribution. Then, it searches the knowledge base for all stored model feature distribution records and finds the most similar model by calculating the Euclidean distance or cosine similarity between feature vectors. This most similar adapted model is selected as the transfer learning base. After the system loads its network structure and parameters, it only needs a few rounds of fine-tuning training with a small amount of early data from the new module to obtain a prediction model suitable for the new module. The entire process typically takes only a few hours to a day, saving years of data accumulation compared to training from scratch.

[0084] After model training, it is lightweighted and embedded into the actual monitoring system, supporting daily automatic feature extraction and real-time inference. The software platform dynamically visualizes the decline trend of the health indicator HI and the prediction results of RUL, and has an early warning function that triggers an alert when the predicted RUL is below 1000 hours or the rate of health decline exceeds a set threshold. The system integrates a closed-loop optimization mechanism, which automatically collects full lifecycle data after actual module failure, and continuously improves the model's prediction accuracy and adaptability in new environments and new batches through sliding window incremental learning and regular model optimization.

[0085] To facilitate understanding of the above technical solutions of the present invention, the following detailed description will be based on the converter system of a large wind farm.

[0086] This wind farm is equipped with 50 wind turbine generators, each with 12 SEMIKRONDANFOSS SKIIP IPM modules in its converter, totaling 600 modules requiring monitoring. Traditional maintenance relies on manual periodic inspections and simple overvoltage and overcurrent protection, often only discovering problems when modules exhibit obvious faults or even smoke and fire, resulting in downtime losses and secondary damage. With this invention, a sampling and conditioning circuit board is first installed in the control cabinet of each converter. The voltage, current, and temperature test points of the 12 IPM modules are connected to the sampling modules via shielded cables, and the data is output to a data acquisition card after multi-stage active filtering. The data acquisition card transmits the digital data to the monitoring workstation in the wind farm's central equipment room via a fiber optic network. Simultaneously, the monitoring software establishes a communication connection with the CAN interface of all IPM modules, enabling synchronous online monitoring of all 600 modules.

[0087] In the initial stage of system operation, all modules were routinely monitored at a baseline sampling frequency of 10kHz. One early morning, the temperature sensor of IPM module No. 5 in Unit 2 detected abnormal fluctuations in junction temperature. The saturation voltage drop characteristic parameter of this module exceeded the preset abnormal threshold three times consecutively within 5 minutes. The monitoring software immediately activated a collaborative monitoring algorithm, increasing the sampling frequency of the abnormal module from 10kHz to 100kHz for finer monitoring. Simultaneously, based on a predefined list of associated modules, the sampling frequencies of adjacent modules No. 4 and No. 6, located in the same bridge arm, were increased to 50kHz, while the sampling frequencies of other normal modules were reduced to 2kHz to conserve system resources. The high-frequency sampled data was rapidly uploaded via the CAN communication module using event-driven PDO messages. Upon detecting the CAN identifier of the faulty PDO, the monitoring software immediately interrupted its current task and parsed the message content. Based on the fault code mapping table, it determined a junction over-temperature fault, classified as a warning. Further analysis using a multi-level fault filtering algorithm revealed that although module No. 4 had not reached the abnormal threshold, its temperature curve also began to show an upward trend. Correlation analysis determined that this bridge arm had a system-level potential problem with an abnormal cooling system. The monitoring interface immediately popped up an alert window to notify the maintenance personnel. At the same time, it recorded the high-frequency waveform data of the past minute into a CSV file for subsequent analysis. It also temporarily lowered the temperature alarm threshold of module 5 from 85 degrees to 75 degrees through an SDO write operation, triggering hardware protection in advance.

[0088] Maintenance personnel arrived at the site and found that the converter's cooling fan was slowing down due to dust blockage. After cleaning the fan, the system returned to normal. After the characteristic parameters of module 5 remained at normal levels for 20 minutes, the monitoring software automatically restored the sampling frequency of all modules to a 10kHz baseline. This fault warning prevented a potential module burnout, saving at least 100,000 yuan in spare parts replacement costs and three days of downtime. More importantly, the system recorded the complete waveform file and processing of this fault event. Using the predictive maintenance module's dynamic time warping algorithm to align multi-source signals, and after wavelet denoising and standardization, extracted multi-dimensional features including switching time extension, junction temperature fluctuation frequency changes, and loss growth trends. Principal component analysis was then used to update the comprehensive health index model for this batch of modules. Based on the historical characteristic sequence of this event, the predictive model estimated that module 5 had approximately 1200 hours of remaining service life before the fault occurred. Comparison with the actual maintenance record of 1850 hours of continuous operation for this module verified the model's predictive validity.

[0089] Three months later, the wind farm introduced a new batch of IPM modules for 10 new turbine units. These new modules had a 20% higher rated current than the old ones, and their switching frequency was increased from 10kHz to 16kHz. Traditionally, it would take at least two years for the new modules to accumulate enough failure data to train the prediction model. However, using the transfer learning mechanism of this invention, the system found the most similar adapted model in terms of electrical characteristics from a cross-model knowledge base. The common degradation patterns learned by this model through pre-training on historical data from the old modules were used as a foundation. Then, the higher-level parameters of the model were fine-tuned using early operating data from the first month of operation of the new modules. During the fine-tuning process, the parameters of the input layer and the first long short-term memory hidden layer were frozen to preserve the ability to extract common features. Only the parameters of the second long short-term memory hidden layer, the fully connected layer, and the output layer were trained, using a small learning rate and an early stopping strategy to prevent overfitting. The training of the new module's adaptation prediction model was completed in just one week, and the model was stored in the knowledge base for use by subsequent modules of the same type. Data verification after six months of actual operation shows that the migration model's prediction error for the remaining service life of the new module is controlled within 15%, which meets the practical requirements. Compared with the traditional method of training from scratch, it significantly shortens the cold start cycle and enables predictive maintenance capabilities to be quickly applied to new equipment.

[0090] like Figure 5 As shown, according to another embodiment of the present invention, an online monitoring and waveform recording system for multiple SEMIKRON DANFOSSSKIIP IPM modules is also provided. This configuration system includes:

[0091] The sampling and conditioning circuit board is used to collect voltage, current, and temperature signals from each IPM module as analog signals, perform multi-stage active filtering on the analog signals, and interact with each IPM module. The input terminal of the sampling and conditioning circuit board is connected to the test point of each IPM module, and the output terminal is connected to the input terminal of the data acquisition card.

[0092] The data acquisition card is used to synchronously sample, amplify, and perform anti-aliasing low-pass filtering on analog signals after multi-stage active filtering to obtain digital data. A cyclic redundancy check mechanism is used for digital data transmission. The input end of the data acquisition card is connected to the output end of the sampling conditioning circuit board through an interface, and the output end is connected to the PC through a network cable or WiFi.

[0093] The PC-based monitoring software dynamically adjusts the monitoring strategies of each IPM module based on its real-time operating status. It interacts with each IPM module using the CANopen high-level communication protocol to receive and display the transmitted digital data in real-time. A buffered writing strategy is employed to record the real-time waveform data as waveform files. Multi-dimensional features are extracted from these waveform files to construct comprehensive health indicators, predict remaining lifespan, and utilize a transfer learning mechanism to enable migration and application across different batches or models of IPM modules. The PC establishes a communication connection with the data acquisition card via Ethernet cable or WiFi.

[0094] The sampling conditioning circuit board includes:

[0095] The sampling module is used to acquire the voltage, current, and temperature signals of each IPM module; this sampling module uses an operational amplifier circuit.

[0096] The filtering module is used to perform multi-stage active filtering on the acquired signal to suppress noise and interference; the input of the filtering module is connected to the output of the sampling module.

[0097] The CAN communication module is used for data interaction with each IPM module. This CAN communication module is connected to the CAN interface of each module through shielded twisted pair cable, and is connected to the PC through an electrically isolated CAN transceiver.

[0098] The PC-based monitoring software includes:

[0099] The monitoring and communication module is used to dynamically adjust the monitoring strategies of each IPM module according to the real-time operating status. It interacts with each IPM module based on the CANopen high-level communication protocol to receive the transmitted digital data and display the real-time waveform. It adopts a buffer writing strategy to record the real-time waveform display data as a waveform file.

[0100] The predictive maintenance module is used to achieve time alignment of multi-source signals in high-frequency data in waveform files using dynamic time warping algorithm, perform signal denoising using wavelet transform algorithm and normalize data using standardization method, extract time domain features, frequency domain features, time-frequency domain features and thermal characteristic features from preprocessed data, use principal component analysis algorithm to fuse multi-dimensional features and reduce dimensions to construct comprehensive health indicators, and use deep long short-term memory network to construct predictive model to predict remaining service life;

[0101] The domain adaptation module is used to take batches of IPM modules with complete remaining useful life tag data as the source domain and newly introduced batches or models of IPM modules lacking remaining useful life tag information as the target domain. The prediction model is pre-trained on the source domain data to extract common temporal features in the health degradation process of IPM modules. The temporal feature sequences of the source and target domains are input into the pre-trained prediction model for domain adaptation training. The prediction model trained by domain adaptation is supervised fine-tuned using early running data in the target domain to obtain the target domain adapted prediction model. When a new IPM module model is added, the target domain adapted prediction model with the most similar feature distribution is selected from the cross-model model knowledge base as the transfer basis, and the network structure and parameters of the target domain adapted prediction model are loaded.

[0102] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for online monitoring and waveform recording of multiple Semikron Danbus Skip IPM modules, characterized in that, include: The voltage, current, and temperature signals of each IPM module are collected as analog signals, multi-stage active filtering is performed on the analog signals, and data is exchanged with each IPM module. The analog signal after multi-stage active filtering is synchronously sampled, amplified, and subjected to anti-aliasing low-pass filtering to obtain digital data, and a cyclic redundancy check mechanism is used for digital data transmission. The monitoring strategy of each IPM module is dynamically adjusted according to the real-time operating status. Data interaction with each IPM module is carried out based on the CANopen high-level communication protocol to realize the reception of the transmitted digital data and the real-time waveform display. A buffer writing strategy is adopted to record the real-time waveform display data as a waveform file. Multidimensional features are extracted from waveform files to construct a comprehensive health index for predicting remaining useful life. The transfer learning mechanism enables the application of this technology across different batches or models of IPM modules.

2. The method for online monitoring and waveform recording of multiple Semikron Danbus Skip IPM modules according to claim 1, characterized in that, The monitoring strategy for dynamically adjusting each IPM module based on its real-time operating status includes: A list of associated modules is predefined for each IPM module, and a baseline sampling frequency is set for all IPM modules; this list of associated modules records other IPM modules that have electrical coupling or physical adjacency with the current IPM module. Real-time monitoring of the characteristic parameters of each IPM module; these characteristic parameters include saturation pressure drop and temperature change rate. When the characteristic parameters of an IPM module exceed the preset abnormal threshold, the IPM module is immediately identified as an abnormal module. The sampling frequency of the abnormal module itself is increased to high frequency mode, and the sampling frequency of the associated modules of the abnormal module is increased to medium frequency mode according to the associated module list. The sampling frequency of normal modules is reduced to low frequency mode to achieve intelligent scheduling of monitoring resources. Once the characteristic parameters of the abnormal module return to normal levels and remain stable for a preset time, the sampling frequency of all IPM modules will be restored to the baseline sampling frequency.

3. The online monitoring and waveform recording method for multiple Semikron Danbus Skip IPM modules according to claim 1, characterized in that, The data interaction with each IPM module based on the CANopen high-level communication protocol includes: An object dictionary mapping table is established for each IPM module instance, mapping the memory addresses of the fault status register, runtime counter, and temperature alarm threshold inside the IPM module to the CANopen object index and sub-index, so as to achieve standardized access to the data inside the module. Synchronization messages are used as a global time reference to synchronize the data acquisition times of all IPM modules; Configure real-time fault status data as an event-driven process data object for transmission; this real-time fault status data configuration is directly triggered by a bit transition in the fault status register inside the IPM module. Configure runtime count and temperature alarm threshold as periodic process data object transfers; The read and write operations of the non-volatile parameters of the IPM module are completed based on the Service Data Object Segmentation Transmission Protocol.

4. The online monitoring and waveform recording method for multiple Semikron Danbus Skip IPM modules according to claim 3, characterized in that, The read / write operations on the non-volatile parameters of the IPM module based on the service data object segmentation transmission protocol include: By sending a service data object write request message, the new operating parameters are written into the non-volatile memory inside the IPM module in several data segments; these operating parameters include the temperature alarm threshold and the fault protection delay. After the IPM module receives all data segments, it verifies the data segments. If the verification is successful, the data segments are stored in non-volatile memory and a successful write response is sent back to the service data object. If an error occurs during transmission, an abort message is sent back and a retransmission is initiated. During the reading process, a service data object read request message is sent, and the parameters in the non-volatile memory are sent back in segments using the IPM module. After all the returned data segments are received, verification is performed to ensure data integrity.

5. The online monitoring and waveform recording method for multiple SEMIKRON DANFOSS SKIIP IPM modules according to claim 3, characterized in that, The data interaction with each IPM module based on the CANopen high-level communication protocol also includes rapid identification of the real-time fault status of the IPM modules, specifically including: When an overcurrent or overtemperature fault occurs inside the IPM module, the state of the dedicated fault pin of the IPM module changes, the fault status register is set, and a pre-mapped process data object message is automatically sent; this process data object message does not require a host request. Upon detecting an identifier from a fault process data object, immediately interrupt the current task and parse the message; Based on a predefined fault code mapping table, specific bytes of the error codes in the process data object are matched with the mapping table to determine the fault level and type, and trigger corresponding protection actions. The fault level includes warning and critical faults, the fault type includes short circuit, drive fault and overheating, and the protection actions include popping up an alert window, recording the fault waveform and issuing a shutdown command.

6. The online monitoring and waveform recording method for multiple SEMIKRON DANFOSS SKIIP IPM modules according to claim 5, characterized in that, The determination of fault level and fault type employs a multi-level fault filtering and dynamic priority parsing algorithm, including: Fault codes are extracted from process data objects, and the fault codes are compared with the validity check matrix to filter out invalid error codes caused by communication interference. The current fault is correlated with recent event logs. Based on the fault code and the correlation analysis results, the fault level and the action to be triggered are determined from the dynamic priority mapping table. The dynamic priority mapping table is dynamically adjusted according to the system operating mode, which includes start-up, steady state, and maintenance mode. When an IPM module reports a fault, the status of its associated modules is queried. If the status of the associated modules is abnormal, it is determined to be a system-level fault and system-level protection is triggered. If the status of the associated modules is normal, local isolation measures for the IPM module are triggered.

7. The method for online monitoring and waveform recording of multiple SEMIKRON DANFOSS SKIIP IPM modules according to claim 1, characterized in that, The digital data transmission using a cyclic redundancy check mechanism includes: Before sending, the data acquisition card, acting as the sender, calculates the first cyclic redundancy check code for the data packet and attaches the first cyclic redundancy check code to the data packet for transmission. After receiving the data packet, the PC-side monitoring software, acting as the receiver, performs a second cyclic redundancy check (CRC) calculation on the data portion of the data packet and determines whether the data transmission is correct by comparing the second CRC with the first CRC. If the second cyclic redundancy check code is equal to the first cyclic redundancy check code, the data is correct; if the second cyclic redundancy check code is not equal to the first cyclic redundancy check code, the data packet is discarded and a retransmission is requested. The method for calculating the cyclic redundancy check code is as follows: shift the data to be checked 32 bits to the left, perform a modulo operation on the fixed generator polynomial, and use the remainder as the check code. The expression for calculating the Cyclic Redundancy Check (CRC) code is: C = (M << 32) mod P; In the formula, C is the cyclic redundancy check code; M is the data message to be checked; and P is the fixed generator polynomial.

8. The method for online monitoring and waveform recording of multiple Semikron Danbus Skip IPM modules according to claim 1, characterized in that, The method of extracting multidimensional features based on waveform files to construct a comprehensive health indicator includes: A dynamic time warping algorithm is used to align the time of multi-source signals in the high-frequency data of the waveform file. Wavelet transform algorithm is used for signal denoising and a standardization method is used to normalize the data to obtain preprocessed data. Time-domain features, frequency-domain features, time-frequency-domain features, and thermal characteristics are extracted from the preprocessed data to serve as multidimensional features. The time-domain features include switching characteristic parameters and waveform statistics; the frequency-domain features include the amplitude of the dominant frequency component extracted by fast Fourier transform of the switching waveform; the time-frequency-domain features include the time-frequency energy distribution features extracted using wavelet basis; and the thermal characteristics include conduction loss, switching loss, junction temperature fluctuation, and mean junction temperature. Principal component analysis algorithm is used to fuse and reduce the dimensionality of multidimensional features to construct a comprehensive health index, and the remaining lifespan is labeled for the preprocessed data.

9. The method for online monitoring and waveform recording of multiple Semikron Danbus Skip IPM modules according to claim 1, characterized in that, The remaining useful life prediction includes: A prediction model is constructed using a deep long short-term memory network. The prediction model includes an input layer, at least two long short-term memory hidden layers, a fully connected layer, and an output layer, which takes a historical feature sequence with continuous time steps as input and outputs the remaining lifetime value of the prediction IPM module. The trained prediction model is embedded into PC-based monitoring software to perform feature extraction and model inference. The PC-based monitoring software displays the decline curve of comprehensive health indicators and the predicted remaining lifespan of the IPM module, and triggers warnings based on preset warning thresholds. When the IPM module fails, the full lifecycle data accumulated from the waveform files of the IPM module is used for incremental learning of the prediction model, and a sliding window time series update strategy is used to periodically train and optimize the prediction model.

10. The method for online monitoring and waveform recording of multiple SEMIKRON DANFOSS SKIIP IPM modules according to claim 9, characterized in that, The migration application on different batches or different models of IPM modules through the transfer learning mechanism includes: The IPM module batches that have accumulated complete remaining service life tag data are used as the source domain, and the newly introduced IPM modules of different batches or different models that lack tag information are used as the target domain. The prediction model is pre-trained on source domain data, and common temporal features in the health degradation process of IPM module are extracted. The output features of the second long short-term memory hidden layer in the pre-trained prediction model are defined as deep shared features. The temporal feature sequences of the source and target domains are input into the pre-trained prediction model. The deep shared features are used to predict the remaining lifetime and calculate the mean squared error loss. The domain distribution difference loss is also calculated. The extended loss function is jointly optimized to complete the domain adaptive training. Supervised fine-tuning of the domain-adaptive prediction model is performed using early running data in the target domain. The network parameters of the input layer and the first long short-term memory hidden layer in the prediction model are frozen. The parameters of the second long short-term memory hidden layer, the fully connected layer and the output layer in the prediction model are fine-tuned and an early stopping strategy is adopted to obtain the target domain-adaptive prediction model. When a new IPM module model is added, the target domain adaptation prediction model with the most similar feature distribution is selected from the cross-model model knowledge base as the migration basis, based on the electrical characteristics, rated parameters and early signal characteristics of the new IPM module. The network structure and parameters of the target domain adaptation prediction model are then loaded.

Images