A variable frequency converter diagnosis method based on LSTM self-encoder multi-model cooperation

Abstract

The application discloses a variable frequency converter diagnosis method based on LSTM self-encoder multi-model cooperation, belongs to the technical field of variable frequency converters, and comprises sensing layer data acquisition and preprocessing, network layer data transmission and task scheduling, platform layer model reasoning and result storage, application layer visualization and feedback optimization, solves the technical problem of effectively identifying time sequence dependent weak abnormal patterns of variable frequency converters under complex working conditions, and realizes adaptive intelligent diagnosis of the variable frequency converter based on deep learning. The application effectively identifies weak abnormal patterns of the variable frequency converter under complex working conditions, greatly reduces the false alarm rate and the missed alarm rate compared with a traditional fixed threshold alarm method, can give an early warning before a fault occurs, saves valuable disposal time for operation and maintenance personnel, continuously improves diagnosis accuracy, greatly reduces the workload of manual piecewise inspection, and improves operation and maintenance efficiency.

Description

Technical Field

[0001] This invention belongs to the field of frequency converter technology, and in particular relates to a frequency converter diagnostic method based on LSTM autoencoder multi-model collaboration. Background Technology

[0002] As the core equipment of industrial drive systems, the real-time monitoring and fault early warning of the operating status of frequency converters are crucial for ensuring production continuity and equipment safety.

[0003] In existing technologies, inverter anomaly detection mainly relies on fixed threshold alarms or fault diagnosis methods based on statistical rules. However, fixed threshold methods are difficult to adapt to the dynamic changes in inverter parameters under different operating conditions, and are prone to generating a large number of false alarms or missed alarms.

[0004] Traditional statistical rule-based methods rely on manual experience to extract features, which cannot effectively identify subtle abnormal patterns under complex operating conditions, especially time-dependent early fault signs.

[0005] In addition, existing solutions are usually only applicable to a single device or brand, lacking cross-manufacturer and cross-model universality, and have not formed a closed-loop mechanism of "detection-prediction-feedback-optimization", making it difficult to continuously improve model accuracy. Summary of the Invention

[0006] The purpose of this invention is to provide a frequency converter diagnosis method based on LSTM autoencoder multi-model collaboration, which solves the technical problem of effectively identifying time-dependent weak anomaly patterns under complex operating conditions of frequency converters and realizing adaptive intelligent diagnosis of frequency converters based on deep learning.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A frequency converter diagnostic method based on LSTM autoencoder multi-model collaboration includes the following steps: The sensor layer collects real-time operating parameters of the frequency converter, performs missing value filling, outlier truncation and standardization processing to obtain standardized time-series window data, and then reports it to the network layer after conversion by the edge gateway protocol; at the same time, it generates anomaly tags based on the device status code. Standardized time-series window data is transmitted from the network layer to the platform layer via 4G / 5G or industrial Ethernet, and the task scheduler asynchronously triggers the judgment task. At the platform layer, standardized time-series window data is input into the LSTM autoencoder model. The reconstruction error is calculated through the encoder-decoder reconstruction algorithm. The comprehensive anomaly score is obtained by combining the output of the classification head and then compared with the threshold to determine the anomaly state. Simultaneously, based on historical data, an LSTM prediction model is used to predict trends. The predicted values ​​are compared with the actual values ​​to calculate the average deviation rate. The average deviation rate is then fused with anomaly indicators to output risk levels and operational recommendations, and the assessment results are stored. The evaluation results are visualized at the application layer, and feedback data with manual annotation is received. The feedback data is fed back to the retraining pool, and the LSTM autoencoder model is periodically retrained using the retraining pool to complete the update of the LSTM autoencoder model.

[0008] Preferably, the missing value imputation adopts the median imputation method, the outlier truncation adopts the 3-sigma principle, and the standardization process adopts StandardScaler normalization based on the mean and variance of the training set. In the perception layer, a sliding window is used to divide historical data into time-series samples, and anomaly labels are generated based on the device status code.

[0009] Preferably, the LSTM autoencoder model includes an encoder, a decoder, and a classification head; The encoder consists of a first LSTM layer with an output dimension of 64, a Dropout layer, and a second LSTM layer with an output dimension of 32 connected sequentially. The decoder consists of a RepeatVector layer, a third LSTM layer with an output dimension of 32, a fourth LSTM layer with an output dimension of 64, and a TimeDistributed fully connected layer connected in sequence. The classification head is responsible for using the 32-dimensional vector generated by the encoder to predict the probability of anomalies.

[0010] Preferably, the formula for calculating the comprehensive anomaly score C5 is as follows: C5=α×p+(1-α)×(1-e -recon_error ); Where p is the anomaly probability output by the classification head of the LSTM autoencoder model, recon_error is the reconstructed mean square error, and α is the preset weight coefficient.

[0011] Preferably, the average deviation rate is specifically calculated using the average absolute percentage deviation, and the specific formula is as follows: Average deviation rate = (1 / N) × (|actual value 1 - predicted value 1| / |actual value 1| + ... + |actual value 1|) N -Predicted value N | / |Actual value N |)×100%; Where N represents the number of prediction steps; Based on the abnormal indicators and average deviation rate, they are mapped to three risk levels: attention, warning, and severe, according to preset rules.

[0012] Preferably, the incremental retraining specifically adopts a small learning rate fine-tuning strategy, extracting at least 500 manually labeled data from the retraining pool and merging them with historical data from the most recent 3 months, and continuing training with the weights of the online model in the current LSTM autoencoder model as the initial parameters. The updated model is only deployed when the anomaly detection accuracy of the new model on the test set is improved by a preset threshold compared to the current online model.

[0013] Preferably, the task scheduler creates an asynchronous assessment task with a period of 60 seconds for each device in the perception layer based on the APScheduler framework, and the resulting task execution progress is pushed to the application layer in real time via WebSocket.

[0014] Preferably, the real-time operating parameters collected by the sensing layer include at least six fields: output current, output voltage, output frequency, DC bus voltage, temperature, and status code. The status code serves as the device status code and is used to generate anomaly tags.

[0015] This invention presents a multi-model collaborative inverter diagnostic method based on LSTM autoencoder, which solves the technical problem of effectively identifying time-dependent weak anomaly patterns in inverters under complex operating conditions and achieving adaptive intelligent diagnosis of inverters based on deep learning. This invention employs a composite loss function jointly trained by an LSTM autoencoder and a classification head, enabling simultaneous learning of the reconstruction accuracy and discriminative ability of time-series features. This effectively identifies weak anomaly patterns in inverters under complex operating conditions. Compared to traditional fixed-threshold alarm methods, it significantly reduces false alarm and false negative rates. Through independent LSTM autoencoder models, it performs multi-step predictions of future key parameters and calculates the average absolute percentage deviation based on actual values. The deviation rate is then fused with anomaly indicators to output a risk level, providing early warnings before faults occur, saving valuable time for maintenance personnel. The front end provides manual annotation boundaries. On the other hand, maintenance personnel can confirm or correct the AI ​​judgment results, and the labeled data is automatically fed back to the retraining pool. The system periodically uses an incremental learning strategy to fine-tune the model, enabling it to adapt to long-term drift such as equipment aging and changes in operating conditions, continuously improving diagnostic accuracy. It can seamlessly connect to inverters from multiple brands such as INVT, Huichuan, and ABB, covering various high-voltage, medium-voltage, and low-voltage models. Based on the APScheduler framework, it realizes timed asynchronous judgment tasks and supports parallel processing of large-scale devices. Combined with WebSocket, the judgment results and alarms are pushed in real time, and the front-end display refreshes in seconds, meeting the stringent real-time requirements of industrial scenarios. Through deep learning models, it automatically completes most of the anomaly identification and trend judgment work, requiring only maintenance personnel to manually label and provide feedback on a small number of edge cases, greatly reducing the workload of manual inspection and improving maintenance efficiency. Attached Figure Description

[0016] Figure 1 This is the main flowchart of the present invention; Figure 2 This is a flowchart of the data acquisition and preprocessing process of the perception layer in this invention; Figure 3 This is a flowchart of network layer data transmission and task scheduling according to the present invention; Figure 4 This is a flowchart of the platform layer model reasoning and result storage, and the application layer visualization and feedback optimization of the present invention. Detailed Implementation

[0017] Depend on Figures 1-4 The inverter diagnostic method based on LSTM autoencoder multi-model collaboration, as shown, includes the following steps: The sensor layer collects real-time operating parameters of the frequency converter, performs missing value filling, outlier truncation and standardization processing to obtain standardized time-series window data, and then reports it to the network layer after conversion by the edge gateway protocol; at the same time, it generates anomaly tags based on the device status code. The real-time operating parameters collected by the sensing layer include at least six fields: output current, output voltage, output frequency, DC bus voltage, temperature, and status code. The status code serves as the device status code and is used to generate anomaly tags.

[0018] The missing value imputation adopts the median imputation method, the outlier truncation adopts the 3-sigma principle, and the standardization process adopts StandardScaler normalization based on the mean and variance of the training set. In the perception layer, a sliding window is used to divide historical data into time-series samples, and anomaly labels are generated based on the device status code.

[0019] In this embodiment, the field inverter reports its operating parameters to the edge gateway (a sensing layer device used to collect the inverter's operating parameters) via the Modbus RTU protocol. The edge gateway can collect data at a frequency of once every 1 second.

[0020] The collected operating parameters include six fields: output current, output voltage, output frequency, DC bus voltage, temperature, and status code. The status code includes 0 for shutdown, 1 for operation, 2 for alarm, and 3 for fault.

[0021] The edge gateway performs preprocessing operations on the collected operating parameters locally, specifically including: first, filling missing values ​​using the median imputation method; then, truncating outliers according to the 3-sigma principle; and finally, using pre-trained StandardScaler parameters (mean and standard deviation) to perform Z-score standardization on the numerical features to obtain standardized time-series window data (e.g., shape [window_size, 5], where window_size = 60). In this embodiment, an original data matrix C0 is constructed, where the matrix has at least 6 columns (i.e., fields of 6 operating parameters, such as output_current), and rows are the actual data.

[0022] The edge gateway transmits standardized window data, along with the device ID and timestamp, to the network layer via 4G / 5G or industrial Ethernet after protocol conversion (e.g., Modbus to MQTT). The network layer then transmits the standardized window data to the platform layer, which can create historical data based on each uploaded standardized window data.

[0023] When training the model at the platform layer, anomaly labels are generated using status codes from historical data. For example, if the status code is 2 (alarm) or 3 (fault), the label is 1 (abnormal); otherwise, it is 0 (normal).

[0024] In this embodiment, during the preprocessing operation, the median imputation method for filling missing values ​​specifically involves running parameters on the five columns of output_current, output_voltage, output_freq, dc_bus_voltage, and temperature, calculating the median_f of the corresponding valid samples (actual data) for each column, and replacing any NaN values ​​in that column with median_f.

[0025] For the column containing the status code, a forward fill method is used, which means filling subsequent missing values ​​with the most recent non-null value to maintain the continuity of the time sequence.

[0026] When truncating outliers using the 3-sigma principle, for each numerical column, the mean μ and standard deviation σ are calculated. Values ​​less than μ-3σ are set as μ-3σ, and values ​​greater than μ+3σ are set as μ+3σ.

[0027] After the above preprocessing operations, the cleaned data matrix C1 can be obtained (including shape [60,5], excluding status, which is only used for anomaly labeling).

[0028] C1 is Z-score standardized using StandardScaler parameters (mean μ_train and standard deviation σ_train for each feature) pre-fitted on the training set: x_scaled=(x-μ_train) / σ_train; Where x is a specific numerical value of a feature, and x_scaled is the standardized value. After standardization, the mean of each feature is 0 and the variance is 1, which is used to eliminate the influence of dimensions.

[0029] In this embodiment, before model training, all normal samples are first extracted from historical data, the mean μ_train and standard deviation σ_train of each numerical feature are calculated, and these parameters are saved for use in the standardization processing of the perception layer and the inference of the platform layer.

[0030] The output of Z-score normalization is normalized window data C2 with shape [60, 5].

[0031] During model training, this embodiment requires dividing the historical data into multiple samples. For example, the window length is set to window_size=60. Starting from the first point of the continuous time series data, 60 consecutive sampling points are taken as the input window X (shape 60×5). Then, the status code of the next sampling point after this window is taken as the label y: if the status is 2 (alarm) or 3 (fault), then y=1 (abnormal); otherwise, y=0 (normal). Then, a step size (default is 1) is slid, repeating until the entire dataset is covered. The generated (X,y) pairs are used for model training.

[0032] The step size can be adjusted according to actual needs. For example, when the step size is 1, the window slides one sampling point at a time, and when the step size is 60, the windows do not overlap.

[0033] The edge gateway reports the standardized window data C2, along with the device ID and timestamp, to the network layer via 4G / 5G or industrial Ethernet after protocol conversion (Modbus to MQTT, etc.). In this embodiment, the perception layer reports once every 60 points to form a standardized window; alternatively, it can report each point in real time, with the platform layer automatically assembling the data into a window.

[0034] Standardized time-series window data is transmitted from the network layer to the platform layer via 4G / 5G or industrial Ethernet, and the task scheduler asynchronously triggers the judgment task. The task scheduler, based on the APScheduler framework, creates an asynchronous assessment task with a period of 60 seconds for each device in the perception layer, and the resulting task execution progress is pushed to the application layer in real time via WebSocket.

[0035] The network layer employs TLS / SSL encrypted transmission and a resume-from-disconnect mechanism to ensure data integrity and security. In this embodiment, a task scheduler is maintained at the platform layer. Based on the APScheduler framework, the task scheduler creates an independent periodic assessment task for each online frequency converter, with a default period of 60 seconds. The task scheduler is configured with key parameters (obtainable from the configuration file loaded at startup), including: assessment period (default 60 seconds), time series window length (default 60 sampling points), anomaly detection threshold (default 0.65), pre-trained model path, and database connection information.

[0036] When each task is triggered, the task scheduler sends an "execution judgment" instruction (carrying the device ID) to the platform layer's model inference service via a message queue (such as RabbitMQ). The task scheduler uses an asynchronous thread pool for execution, supporting parallel processing of high-concurrency devices.

[0037] The scheduler will push the execution progress (0% to 100%) of each task to the front-end application layer in real time via the WebSocket protocol, making it easy for operations and maintenance personnel to monitor.

[0038] At the platform layer, standardized time-series window data is input into the LSTM autoencoder model. The reconstruction error is calculated through the encoder-decoder reconstruction algorithm. The comprehensive anomaly score is obtained by combining the output of the classification head and then compared with the threshold to determine the anomaly state. Simultaneously, based on historical data, an LSTM prediction model is used to predict trends. The predicted values ​​are compared with the actual values ​​to calculate the average deviation rate. The average deviation rate is then fused with anomaly indicators to output risk levels and operational recommendations, and the assessment results are stored. The LSTM autoencoder model includes an encoder, a decoder, and a classification head; The encoder consists of a first LSTM layer with an output dimension of 64, a Dropout layer, and a second LSTM layer with an output dimension of 32 connected sequentially. The decoder consists of a RepeatVector layer, a third LSTM layer with an output dimension of 32, a fourth LSTM layer with an output dimension of 64, and a TimeDistributed fully connected layer connected in sequence. The classification head is responsible for using the 32-dimensional vector generated by the encoder to predict the probability of anomalies.

[0039] The formula for calculating the comprehensive anomaly score C5 is as follows: C5=α×p+(1-α)×(1-e -recon_error ); Where p is the anomaly probability output by the classification head of the LSTM autoencoder model, recon_error is the reconstruction mean square error, and α is a preset weight coefficient. In this embodiment, when the reconstruction error is very small, (1-e -recon_error When the reconstruction error is large, this item approaches 1; when the reconstruction error is large, this item approaches 0. The overall score C5 is between 0 and 1, and the higher the score, the more likely it is to be an anomaly. The value of α is 0.5.

[0040] In this embodiment, the reconstruction loss weight is 0.7 and the classification loss weight is 0.3 during training; during inference, α = 0.5. These two weights differ: the training weights are used to balance the direction of model parameter optimization, while the inference weight is used to fuse the two heterogeneous signals; both can be set independently.

[0041] The average deviation rate is specifically calculated using the average absolute percentage deviation, and the specific formula is as follows: Average deviation rate = (1 / N) × (|actual value 1 - predicted value 1| / |actual value 1| + ... + |actual value 1|) N -Predicted value N | / |Actual value N |)×100%; Where N represents the number of prediction steps; in this embodiment, if the actual value at a certain moment is 0, then the absolute error |actual value - predicted value| is used instead, that is, the contribution of this item is |actual value - predicted value|.

[0042] Based on the abnormal indicators and average deviation rate, they are mapped to three risk levels: attention, warning, and severe, according to preset rules.

[0043] In this embodiment, the encoder's task is to compress a 60×5 temporal window into a compact 32-dimensional vector, which is called the encoded vector or latent representation. The encoder internally uses a two-layer LSTM network: The first LSTM layer contains 64 memory units. It processes data at 60 time points sequentially, outputting an intermediate result at each time point. Therefore, the first layer outputs a sequence of 60 time steps, with each time step corresponding to 64 values.

[0044] This is followed by a Dropout layer, which randomly ignores 20% of the neuron outputs during training to prevent the model from overfitting.

[0045] The second LSTM layer contains 32 memory units. It only outputs the result of the last time step, which is the encoded vector (32 values).

[0046] In this embodiment, each LSTM layer uses the default tanh activation function and sigmoid loop activation function. The tanh activation function and sigmoid loop activation function are existing technologies and will not be described in detail.

[0047] The decoder's task is to receive the 32-dimensional encoded vector generated by the encoder and attempt to reconstruct the original 60×5 sequence. The decoder's working steps are as follows: First, the 32-dimensional encoding vector is copied 60 times and stacked into a 60×32 matrix. This ensures that the decoder has the same starting input at every time step.

[0048] Then, it goes through two LSTM layers: the first layer has 32 memory units and outputs the intermediate results at each time step; the second layer has 64 memory units and outputs richer features at each time step.

[0049] Finally, a fully connected layer (Dense layer) is used to map the 64-dimensional features at each time step back to 5 dimensions, that is, 5 numerical features at each time point.

[0050] The decoder ultimately outputs a 60×5 reconstructed sequence.

[0051] The classification head is an additional branch that directly uses the 32-dimensional vector generated by the encoder to predict the anomaly probability. The classification head consists of two fully connected layers.

[0052] The first layer maps 32 dimensions to 32 dimensions and introduces nonlinearity using the ReLU activation function.

[0053] The second layer maps 32 dimensions to a single numerical value, and then uses the Sigmoid function to compress the value to between 0 and 1, outputting an anomaly probability (0 indicates that the model considers it normal, and 1 indicates that it considers it abnormal).

[0054] During training, a batch of historical data is used, which contains a large number of normal windows (label y=0) and a small number of abnormal windows (label y=1, i.e., the next status code after the window is 2 or 3). For each training sample (input window X, label y), the model simultaneously performs encoding, decoding, and classification, and then calculates two losses.

[0055] The reconstruction loss is calculated as follows: the mean squared error between the input window X and the reconstruction window output by the decoder is calculated. Specifically, the values ​​at each corresponding position in X and the reconstruction window are subtracted, squared, summed, and then the average value is taken. A smaller reconstruction loss indicates a stronger ability of the LSTM autoencoder model to reconstruct the normal pattern.

[0056] The classification loss is calculated as follows: the binary cross-entropy between the anomaly probability p output by the classification head and the true label y is calculated. The closer p and y are, the smaller the classification loss.

[0057] The final total loss is: Total Loss = 0.7 × Reconstruction Loss + 0.3 × Classification Loss. 0.7 and 0.3 are the default weights for reconstruction loss and classification loss, respectively. Setting them to 0.7 and 0.3 indicates that the model training focuses more on learning the temporal patterns of normal data (reconstruction ability) while also taking into account the accuracy of anomaly classification.

[0058] In this embodiment, training can use the Adam optimizer, with a learning rate of 0.001, 100 epochs, and each batch containing 64 samples. After training, the model is saved as a file.

[0059] When the task scheduler issues an assessment command for a certain device (e.g., GD5000-001), the platform layer executes the following process: The standardized data (i.e., the C2 data blocks reported and stored by the perception layer) of the device's most recent 60 sampling points are retrieved from the time-series database to obtain the standardized window data. If there are fewer than 60 data points, the process waits until the next cycle before proceeding.

[0060] The standardized window data is input into the trained LSTM autoencoder model, and the model outputs two results: the reconstructed window (the restored 60×5 sequence) and the anomaly probability p (a value between 0 and 1 given by the classification head).

[0061] The reconstruction error, recon_error, is obtained by subtracting, squaring, and summing the values ​​at each corresponding position in the original input window and the reconstruction window, and then averaging the results (i.e., the mean squared error algorithm, MSE). The larger the recon_error, the greater the difference between the current input sequence pattern and the normal pattern.

[0062] A comprehensive anomaly score C5 is calculated based on the anomaly probability p output by the classification head and the reconstruction error.

[0063] The overall score C5 is compared with a preset threshold (0.65). If the overall score is greater than 0.65, the current window is determined to be "abnormal"; otherwise, it is determined to be "normal".

[0064] In this embodiment, the platform layer also deploys an independent LSTM prediction model to predict the changing trends of key parameters over a future period. This prediction model uses historical output current data from the past 7 days (168 points, one sampling point per hour) to predict the output current value for the next 24 hours (24 points). The LSTM prediction model employs an encoder-decoder structure, and its training method is similar to that of the LSTM autoencoder model, but the objective is different: the LSTM prediction model learns how to predict future sequences based on historical sequences.

[0065] Specifically, the LSTM prediction model adopts an encoder-decoder architecture. The encoder consists of two LSTM layers. The first layer has 64 units (return_sequences=True), and the second layer has 32 units (return_sequences=False), encoding 168 historical current values into a 32-dimensional vector. The decoder repeats this vector 24 times, and then passes it through two LSTM layers (the first layer has 32 units with return_sequences=True, and the second layer has 64 units with return_sequences=True) and a fully connected layer (outputting 1 dimension) to obtain 24 predicted values. During training, the historical current sequence is used as the input, and the corresponding 24 future real current values are used as labels. The loss function is the mean squared error (MSE). The optimizer uses Adam with a learning rate of 0.001. return_sequences=True means outputting the results of each time step, and return_sequences=False means only outputting the results of the last time step.

[0066] After the actual current values for the next 24 hours are collected, the actual values are compared with the predicted values, and the mean absolute percentage error (MAPE), that is, the average deviation rate, is calculated.

[0067] When fusing the average deviation rate with the anomaly flag, the preset rules are as follows: If the anomaly flag is 1 and MAPE > 15%, the risk level is "severe", and the suggestion is "immediately stop for maintenance".

[0068] If the anomaly flag is 1 or MAPE > 15% (either one is satisfied), the risk level is "warning", and the suggestion is "arrange an inspection within this week".

[0069] If the anomaly flag is 0 and 10% < MAPE ≤ 15%, the risk level is "attention", and the suggestion is "strengthen monitoring and pay attention to trends".

[0070] In other cases, the risk level is "normal", and the suggestion is "no intervention required".

[0071] In this embodiment, after each judgment is completed, the platform layer will write the following information into the "judgment result table" of the MySQL database: device number, judgment timestamp, comprehensive anomaly score, reconstruction error, judgment result ("normal" or "abnormal"), risk level (severe / warning / attention / normal), and operation and maintenance suggestion text.

[0072] At the same time, the platform layer can also publish a notice of "new judgment result" through a message queue (RabbitMQ) for the application layer to subscribe to and refresh the interface in real time.

[0073] The evaluation results are visualized at the application layer, and feedback data with manual annotation is received. The feedback data is fed back to the retraining pool, and the LSTM autoencoder model is periodically retrained using the retraining pool to complete the update of the LSTM autoencoder model.

[0074] The incremental retraining specifically adopts a small learning rate fine-tuning strategy. At least 500 manually labeled data points are extracted from the retraining pool and merged with historical data from the most recent 3 months. The weights of the online model in the current LSTM autoencoder model are used as initial parameters for continued training. The updated model is only deployed when the anomaly detection accuracy of the new model on the test set is more than 1% higher than that of the current online model.

[0075] In this embodiment, the application layer can build front-end pages based on the Vue3 and ECharts frameworks and interact with the platform layer through REST APIs. For example: When the page loads, the API GET / api / devices is called to obtain the latest assessment status of all frequency converters, and then each device is displayed in the form of a card: normal devices are displayed as green cards, and abnormal devices are displayed as red cards.

[0076] After clicking on a device card, the API GET / api / anomalies?device_id=xxx is called to retrieve the historical anomaly record list for that device. The list displays the time of each assessment, anomaly score, reconstruction error, AI assessment result, and manual annotation status.

[0077] Clicking on an abnormal record will bring up a details window, displaying a comparison chart of the original time series curve and the model reconstruction curve, as well as a line graph comparing the predicted current value and the actual value.

[0078] In addition, a long WebSocket connection is maintained between the front-end and the platform layer. When the platform layer generates new analysis results or alarms, it will actively push them to the front-end. The front-end will immediately update the card color and abnormal list of the corresponding device, achieving a second-level refresh.

[0079] When providing manual annotation feedback, each anomaly record on the front end has a "Manual Annotation" button next to it. For example, when the operations and maintenance personnel (experts) click the button, a dialog box will pop up with three options: "Normal (False Alarm)," indicating that the anomaly determined by the model is actually a normal situation; "Anomaly (Correct)," indicating that the anomaly determined by the model is a real fault; and "Anomaly (Missed Alarm)," indicating that the normal situation determined by the model is actually a missed anomaly.

[0080] Operations personnel can also add comments. After submission, the front-end calls the API POST / api / label to send the record number, manual label, and comments to the platform layer.

[0081] After receiving the annotation request, the platform layer updates the corresponding judgment entry in the "Judgment Result Table" of the MySQL database, and inserts the original feature sequence (standardized window data) of the judgment and the manual label into the "Retraining Pool" table for subsequent incremental training.

[0082] This embodiment also sets up a scheduled task that runs once a week (e.g., every Sunday at 2 AM). This task performs the following operations: Read all manually labeled data that have not yet been used for training from the retraining pool, with a minimum of 500 data entries required.

[0083] Extract the most recent 3 months of normal operation data from the historical database (data with status code 1 and not marked as abnormal).

[0084] The two sets of data were combined and divided into training, validation and test sets in a ratio of 7:2:1.

[0085] Load the weights of the currently used LSTM autoencoder model as initial parameters. Set a small learning rate (e.g., 1e-5) and continue training on the training set for 20 epochs.

[0086] Evaluate the anomaly detection accuracy of the new model on the test set. If the accuracy is more than 1% higher than the current online model, save the new model file and increment the version number (e.g., change from version001 to version002).

[0087] To update the online model using a blue-green deployment approach: Place the new model file in the version directory of the TensorFlowServing service (e.g., / models / inverter_anomaly / 002 / ), modify the service configuration file to point the latest version policy to version 2, and then send a hot reload request. The service will automatically switch to the new model, and the entire process will not interrupt the inference service.

[0088] The old version model is retained for 7 days so that it can be rolled back quickly in case of problems.

[0089] Through manual annotation and incremental retraining, the model can continuously adapt to long-term drift such as aging of frequency converter equipment and changes in operating conditions, thereby continuously improving diagnostic accuracy.

[0090] This invention presents a multi-model collaborative inverter diagnostic method based on LSTM autoencoder, which solves the technical problem of effectively identifying time-dependent weak anomaly patterns in inverters under complex operating conditions and achieving adaptive intelligent diagnosis of inverters based on deep learning. This invention employs a composite loss function jointly trained by an LSTM autoencoder and a classification head, enabling simultaneous learning of the reconstruction accuracy and discriminative ability of time-series features. This effectively identifies weak anomaly patterns in inverters under complex operating conditions. Compared to traditional fixed-threshold alarm methods, it significantly reduces false alarm and false negative rates. Through independent LSTM autoencoder models, it performs multi-step predictions of future key parameters and calculates the average absolute percentage deviation based on actual values. The deviation rate is then fused with anomaly indicators to output a risk level, providing early warnings before faults occur, saving valuable time for maintenance personnel. The front end provides manual annotation boundaries. On the other hand, maintenance personnel can confirm or correct the AI ​​judgment results, and the labeled data is automatically fed back to the retraining pool. The system periodically uses an incremental learning strategy to fine-tune the model, enabling it to adapt to long-term drift such as equipment aging and changes in operating conditions, continuously improving diagnostic accuracy. It can seamlessly connect to inverters from multiple brands such as INVT, Huichuan, and ABB, covering various high-voltage, medium-voltage, and low-voltage models. Based on the APScheduler framework, it realizes timed asynchronous judgment tasks and supports parallel processing of large-scale devices. Combined with WebSocket, the judgment results and alarms are pushed in real time, and the front-end display refreshes in seconds, meeting the stringent real-time requirements of industrial scenarios. Through deep learning models, it automatically completes most of the anomaly identification and trend judgment work, requiring only maintenance personnel to manually label and provide feedback on a small number of edge cases, greatly reducing the workload of manual inspection and improving maintenance efficiency.

Claims

1. A frequency converter diagnostic method based on LSTM autoencoder multi-model collaboration, characterized in that: The steps include the following: The sensor layer collects real-time operating parameters of the frequency converter, performs missing value filling, outlier truncation and standardization processing to obtain standardized time-series window data, and then reports it to the network layer after conversion by the edge gateway protocol; at the same time, it generates anomaly tags based on the device status code. Standardized time-series window data is transmitted from the network layer to the platform layer via 4G / 5G or industrial Ethernet, and the task scheduler asynchronously triggers the judgment task. At the platform layer, standardized time-series window data is input into the LSTM autoencoder model. The reconstruction error is calculated through the encoder-decoder reconstruction algorithm. The comprehensive anomaly score is obtained by combining the output of the classification head and then compared with the threshold to determine the anomaly state. Simultaneously, based on historical data, an LSTM prediction model is used to predict trends. The predicted values ​​are compared with the actual values ​​to calculate the average deviation rate. The average deviation rate is then fused with anomaly indicators to output risk levels and operational recommendations, and the assessment results are stored. The evaluation results are visualized at the application layer, and feedback data with manual annotation is received. The feedback data is fed back to the retraining pool, and the LSTM autoencoder model is periodically retrained using the retraining pool to complete the update of the LSTM autoencoder model.

2. The inverter diagnostic method based on LSTM autoencoder multi-model collaboration as described in claim 1, characterized in that: The missing value imputation adopts the median imputation method, the outlier truncation adopts the 3-sigma principle, and the standardization process adopts StandardScaler normalization based on the mean and variance of the training set. In the perception layer, a sliding window is used to divide historical data into time-series samples, and anomaly labels are generated based on the device status code.

3. The inverter diagnostic method based on LSTM autoencoder multi-model collaboration as described in claim 1, characterized in that: The LSTM autoencoder model includes an encoder, a decoder, and a classification head; The encoder consists of a first LSTM layer with an output dimension of 64, a Dropout layer, and a second LSTM layer with an output dimension of 32 connected sequentially. The decoder consists of a RepeatVector layer, a third LSTM layer with an output dimension of 32, a fourth LSTM layer with an output dimension of 64, and a TimeDistributed fully connected layer connected in sequence. The classification head is responsible for using the 32-dimensional vector generated by the encoder to predict the probability of anomalies.

4. The inverter diagnostic method based on LSTM autoencoder multi-model collaboration as described in claim 3, characterized in that: The formula for calculating the comprehensive anomaly score C5 is as follows: C5 = α × p + (1 - α) × (1 - e -recon_error ); Where p is the anomaly probability output by the classification head of the LSTM autoencoder model, recon_error is the reconstructed mean square error, and α is the preset weight coefficient.

5. The inverter diagnostic method based on LSTM autoencoder multi-model collaboration as described in claim 1, characterized in that: The average deviation rate is specifically calculated using the average absolute percentage deviation, and the specific formula is as follows: Average deviation rate = (1 / N) × (|actual value 1 - predicted value 1| / |actual value 1| + ... + |actual value 1|) N -Predicted value N | / |Actual value N |)×100%; Where N represents the number of prediction steps; Based on the abnormal indicators and average deviation rate, they are mapped to three risk levels: attention, warning, and severe, according to preset rules.

6. The inverter diagnostic method based on LSTM autoencoder multi-model collaboration as described in claim 1, characterized in that: The incremental retraining specifically adopts a small learning rate fine-tuning strategy. At least 500 manually labeled data points are extracted from the retraining pool and merged with historical data from the most recent 3 months. The weights of the online model in the current LSTM autoencoder model are used as initial parameters for continued training. The updated model is only deployed when the anomaly detection accuracy of the new model on the test set is improved by a preset threshold compared to the current online model.

7. The inverter diagnostic method based on LSTM autoencoder multi-model collaboration as described in claim 1, characterized in that: The task scheduler, based on the APScheduler framework, creates an asynchronous assessment task with a period of 60 seconds for each device in the perception layer, and the resulting task execution progress is pushed to the application layer in real time via WebSocket.

8. The inverter diagnostic method based on LSTM autoencoder multi-model collaboration as described in claim 1, characterized in that: The real-time operating parameters collected by the sensing layer include at least six fields: output current, output voltage, output frequency, DC bus voltage, temperature, and status code. The status code serves as the device status code and is used to generate anomaly tags.

Images