Abnormity detection and intelligent error correction system and method for machining process of spark machine

Abstract

The invention relates to the technical field of data processing, and discloses a spark machine machining process anomaly detection and intelligent error correction system and method.The spark machine machining process anomaly detection and intelligent error correction system comprises a sensor module used for collecting multi-source time sequence data in the machining process in real time; the signal processing module is used for carrying out preprocessing and feature extraction on the multi-source time sequence data to obtain a time domain feature set and a frequency domain feature set; the anomaly detection module is used for outputting state identifiers representing the machining state according to the time domain feature set and the frequency domain feature set, and the state identifiers comprise a normal state, an early warning state and an alarm state; the intelligent decision-making module is configured to call a preset error correction strategy library according to the state identifier and generate a corresponding control instruction; and the execution control module is used for executing the control instruction so as to adjust the machining parameters of the spark machine or perform start-stop control on the machining process, so that real-time anomaly detection and self-adaptive error correction in the machining process are realized, the machining precision, the equipment safety and the production efficiency are improved, and the manual intervention and the maintenance cost are reduced.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a system and method for detecting and intelligently correcting anomalies in the EDM machining process. Background Technology

[0002] A spark erosion machine, also known as an electrical discharge machining (EDM) machine tool, is a specialized machine tool that uses the electro-erosion phenomenon generated by pulsed spark discharge between two electrodes immersed in a working fluid to remove conductive materials. It employs a technique that uses the principle of electrical discharge corrosion to precisely machine conductive materials and is widely used in industries such as mold manufacturing and aerospace.

[0003] However, during the processing, factors such as unstable discharge state, electrode wear, and abnormal working fluid can easily lead to a decline in processing quality or even equipment damage. Traditional EDM monitoring systems are mostly based on single sensor signals and simple threshold judgments, which are difficult to comprehensively reflect the processing status and lack effective intelligent error correction mechanisms. They often rely on manual intervention, which is inefficient and prone to errors. In addition, although some systems have attempted to monitor the status using multi-sensor data, they still have shortcomings in terms of the comprehensiveness of feature extraction, the accuracy of anomaly identification, and the intelligence of error correction strategies. Summary of the Invention

[0004] To improve the comprehensiveness, accuracy, and intelligence of the EDM monitoring system, this application provides an EDM process anomaly detection and intelligent error correction system, employing the following technical solution: A fault detection and intelligent error correction system for EDM machining process includes: The sensor module is installed in the discharge circuit, electrode system and working fluid system of the EDM machine to collect multi-source time-series data in real time during the processing. The multi-source time-series data includes discharge voltage, discharge current, electrode vibration signal, electrode temperature signal, working fluid flow rate and working fluid pressure. The signal processing module is communicatively connected to the sensor module and is used to preprocess, feature association model, and feature extraction of the multi-source time-series data to obtain a time-domain feature set, a frequency-domain feature set, and an association-fused feature set. An anomaly detection module, which is communicatively connected to the signal processing module, includes a processing stage identification unit and a multimodal temporal anomaly detection model. It is used to output a status identifier representing the processing state based on the time-domain feature set, frequency-domain feature set, and associated fusion feature set. The status identifier includes normal state, early warning state, and alarm state. The intelligent decision-making module, which is communicatively connected to the anomaly detection module, is configured to call a preset three-level mapping error correction strategy library and generate corresponding control commands based on the state identifier and the corresponding anomaly feature combination. The execution control module is communicatively connected to the intelligent decision-making module and the CNC system of the EDM machine, and is used to execute the control commands and verify and provide feedback on the error correction effect.

[0005] By adopting the above technical solutions, multi-sensor data fusion and intelligent analysis can be used to achieve real-time anomaly detection and adaptive error correction in the EDM machining process, thereby improving machining accuracy, equipment safety and production efficiency, and reducing manual intervention and maintenance costs.

[0006] Optionally, the signal processing module includes a feature correlation analysis unit and a feature importance evaluation unit, specifically configured as follows: Preprocessing unit: Employs a differentiated adaptive filtering algorithm to denoise multi-source time-series data; Feature association analysis unit: Calculates the association degree of multi-source features based on mutual information entropy, establishes an association matrix, and performs weighted fusion of strongly associated features to obtain associated fused features; Feature extraction unit: calculates time-domain features and frequency-domain features. The time-domain features include mean, variance, peak value and waveform factor. The frequency-domain features include the main frequency amplitude and spectral centroid obtained after performing a fast Fourier transform on the preprocessed signal. Feature importance evaluation unit: The ReliefF algorithm is used to calculate the discriminative power of each feature, and redundant features with discriminative power below a specified value are removed, while the core feature set is retained.

[0007] Optionally, the anomaly detection module includes a multimodal temporal anomaly detection model; The multimodal temporal anomaly detection model takes the temporal feature vector composed of the temporal feature set, frequency feature set, and correlation fusion feature set as input, and includes a bidirectional gated recurrent unit network based on attention mechanism, a sample verification unit, and a model update unit. The bidirectional gated recurrent unit network is used to encode the temporal feature vector and calculate the hidden state at each time step; the output layer of the multimodal temporal anomaly detection model maps the final hidden state to the state identifier. The sample verification unit is configured as follows: When the status identifier is in a normal state, the processing procedure corresponding to the current time series feature vector is confirmed to be "true normal working condition" through multi-dimensional consistency verification. When the status identifier is a warning or alarm status, record the time-series feature vector and the corresponding processing scenario information. After manual review or fault tracing verification, label its true abnormal type. The model update unit is configured as follows: Regularly collect real normal samples confirmed by the sample verification unit and labeled real abnormal samples to construct an incremental training set; based on this incremental training set, trigger the incremental learning process of the model and dynamically adjust the parameters of the bidirectional gated recurrent unit network.

[0008] By adopting the above technical solutions, the incremental learning mechanism enables the system to update online and adapt to the operating condition drift caused by equipment wear, material changes, etc., which to a certain extent reduces the occurrence of the traditional fixed model's "performance degradation begins as soon as it goes online", and improves the long-term stability and practicality of the system.

[0009] Optionally, the computation process of the attention-based bidirectional gated recurrent unit network includes: The forward computation layer processes temporal feature vectors in chronological order to obtain the forward hidden state sequence. ; The backward computation layer processes the temporal feature vectors in reverse chronological order to obtain the backward hidden state sequence. ; Each time step The forward hidden state and the backward hidden state are concatenated to obtain the hidden state containing complete context information. ; Calculate the hidden state at each time step attention weights The hidden states at all time steps are summed in weight according to their attention weights to obtain the final context vector. , , This represents the total number of time steps of the processed temporal feature vectors; The final context vector It is fed into the output layer for classification.

[0010] By adopting the above technical solutions, the attention mechanism can provide the basis for the model to make decisions, that is, to visually show which features at which time points caused the "warning" or "alarm". This provides operators with valuable diagnostic information, making it easier to quickly locate the root cause of the problem. It realizes the functional upgrade from detection to auxiliary diagnosis and has a higher degree of intelligence.

[0011] Optionally, the attention weights The calculation methods include: Where j represents the loop counter variable, indicating that each time step from 1 to T is traversed during the summation process. Represents the hidden state at time step j. This is the attention scoring function.

[0012] Optionally, the attention scoring function is implemented using a feedforward neural network, which includes fully connected layers and non-linear activation functions. ; The calculation formula for the attention scoring function includes: in, It is a trainable weight matrix. It is a trainable weight vector used to compress the mapped result into a scalar score. It is the hyperbolic tangent activation function, used to introduce nonlinearity. It is a trainable bias vector.

[0013] By adopting the above technical solution, the trainable feedforward neural network can learn the dynamic weight allocation strategy of different signal characteristics at different times during the EDM process through data. For example, the current stability is most important at the beginning of the EDM discharge, while the frequency characteristics of electrode vibration are more discriminative in the middle of the process. This dynamic and adaptive feature importance assessment capability at different stages is one of the core advantages of this solution compared with traditional fixed threshold or static models.

[0014] Optionally, the anomaly detection module includes an output module, and the output module's status identifier output rule is as follows: Output normal status when all features are within the normal range; A warning status is output when at least one feature exceeds the normal range but does not reach the alarm threshold. An alarm status is output when any feature reaches or exceeds the alarm threshold.

[0015] Optionally, it may also include a display alarm module and a data storage module; The display alarm module is used to display the processing status, feature data, status identifier and error correction action in real time, and to provide audible and visual alarms in the early warning and alarm states. The data storage module is used to store multi-source time-series data, feature data, historical records of status identifiers, and error correction logs, and supports data backtracking and analysis.

[0016] Secondly, this application provides a method for abnormal detection and intelligent error correction in the EDM machining process, which adopts the following technical solution: A method for detecting and intelligently correcting anomalies in the EDM machining process, applicable to any of the EDM machining process anomaly detection and intelligent error correction systems described above.

[0017] In summary, this application includes at least one of the following beneficial technical effects: This application utilizes a sensor module to collect real-time multi-source time-series data, including discharge voltage, discharge current, electrode vibration signals, working fluid flow rate, and working fluid pressure. A signal processing module performs filtering, noise reduction, and normalization preprocessing on the data, extracting time-domain features (mean, variance, peak value, waveform factor) and frequency-domain features (dominant frequency amplitude, spectral centroid). An anomaly detection module analyzes the feature set using a support vector machine model, outputting normal, warning, or alarm status identifiers. An intelligent decision-making module selects the appropriate strategy from the error correction strategy library based on the status identifiers, generating control commands. The execution control module executes these commands through the CNC system interface, adjusting machining parameters or starting / stopping machining. This achieves multi-sensor data fusion and intelligent analysis, enabling real-time anomaly detection and adaptive error correction during the EDM machining process. This improves machining accuracy, equipment safety, and production efficiency, while reducing manual intervention and maintenance costs. Attached Figure Description

[0018] Figure 1 This is a structural block diagram of the EDM machining process anomaly detection and intelligent error correction system in this application. Detailed Implementation

[0019] The embodiments of this application are described in detail below, and examples of the embodiments are shown in the accompanying drawings.

[0020] In the description of this specification, the references to "certain embodiments," "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples" refer to specific features, structures, materials, or characteristics described in connection with the described embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0021] The following detailed description of the EDM machining process anomaly detection and intelligent error correction system of this application, in conjunction with the accompanying drawings and specific embodiments, is provided. This embodiment uses a precision EDM-800 model as the adaptable object and is applied to the precision mold machining scenario of aerospace parts. The machining material is SKD11 mold steel, and the machining accuracy requirement is ±0.01mm. This system realizes real-time monitoring and intelligent optimization of the machining process.

[0022] This application discloses an abnormality detection and intelligent error correction system for EDM machining process, referring to... Figure 1It includes a sensor module, a signal processing module, an anomaly detection module, an intelligent decision-making module, and an execution control module. Each module communicates via industrial Ethernet (Profinet protocol). The sensor module and the execution control module communicate bidirectionally with the EDM CNC system (FANUC0i-MF) through the IO-Link interface. The data storage module uses an SQL Server 2019 database, supporting the writing and querying of 1,000 data entries per second. The display and alarm module uses a 10-inch industrial touch screen with a resolution of 1280×800 to achieve data visualization and operation interaction.

[0023] Specifically, the sensor modules are deployed according to the key system zones of the EDM machine, and the specific selection and installation methods are as follows: Discharge circuit sensors: A Hall current sensor of model CYH-100A (measurement range 0-100A, accuracy ±0.5%) and a voltage sensor of model CYV-10kV (measurement range 0-10kV, accuracy ±0.2%) are selected and installed at the output terminal of the discharge power supply of the EDM machine. The sampling frequency is set to 10kHz to collect the time sequence data of discharge current and discharge voltage in real time. Electrode system sensors: A triaxial accelerometer of model SDV-600 (measuring range ±50g, sensitivity 100mV / g) is used, which is attached to the electrode clamp to collect electrode vibration signals (sampling frequency 5kHz); a platinum resistance temperature sensor of model PT100 (measuring range 0-200℃, accuracy ±0.1℃) is selected and embedded inside the electrode rod to monitor the electrode temperature in real time. Working fluid system sensors: A turbine flow meter of model LWGY-15 (measuring range 0.1-1m³ / h, accuracy ±1%) is used and installed in the working fluid inlet pipe to collect the working fluid flow rate; a pressure sensor of model MPX5700 (measuring range 0-1MPa, accuracy ±0.3%) is selected and installed at the working fluid circulation pump outlet to monitor the working fluid pressure. Both sensors have a sampling frequency of 1kHz. The output signals of all sensors are converted into 4-20mA standard signals by a signal conditioner (model: ADAM-3017) and then transmitted to the signal processing module.

[0024] The signal processing module is implemented based on an industrial control computer (CPU: Intel Core i7-12700, 32GB RAM, 1TB SSD), equipped with a Python 3.9 programming environment, and uses NumPy and SciPy libraries to complete data processing. The specific process is as follows: S1. Perform differential preprocessing: For the discharge voltage / current signal, a 3-level wavelet decomposition is performed using the db4 wavelet basis, and an adaptive threshold is calculated based on the signal-to-noise ratio (SNR) (threshold formula: ,in (where N is the signal length and the noise standard deviation is denoted as N), the wavelet coefficients are thresholded and the signal is reconstructed to achieve pulse edge preservation and noise suppression. The signal-to-noise ratio of the processed signal is increased from the original 25dB to 50dB. For the electrode vibration signal, ±3 is first removed by amplitude limiting filtering. To remove pulse interference, a fourth-order Butterworth bandpass filter with a 50-500Hz passband is used to retain the characteristic frequency components caused by electrode wear and eccentricity, while filtering out high-frequency electromagnetic interference and low-frequency mechanical vibration noise. For the working fluid flow / pressure signal, a first-order hysteresis low-pass filter (time constant) is used. The formula is This smooths out instantaneous fluctuations in flow and pressure. Finally, normalization is performed, and the interval scaling method is used to map all preprocessed signals to the [0,1] interval.

[0025] S2. Perform feature association analysis: The correlation degree of multi-source features is calculated based on mutual information entropy (MI). The formula for calculating mutual information entropy is: ,in , The information entropy of features X and Y are respectively. For joint information entropy; Establish a correlation matrix of "discharge parameters - electrode state - working fluid state", where the matrix elements are the mutual information entropy values ​​of the corresponding features. For example, the MI of discharge current variance and electrode vibration peak value is 0.78 (strong correlation), and the MI of working fluid flow rate and electrode temperature is 0.65 (moderate correlation). Strongly correlated features with mutual information entropy ≥ 0.6 are weighted and fused, with weights assigned based on feature importance coefficients (e.g., discharge current variance weight 0.5, electrode vibration peak weight 0.5), resulting in two types of correlated fused features: "discharge-electrode synergistic features" and "electrode-working fluid compatibility features".

[0026] S3. Perform feature extraction: For time-domain characteristics, the mean (reflecting the average level of the signal), variance (reflecting the stability of the signal), peak value (reflecting the extreme value of the signal), and waveform factor (the ratio of peak value to effective value, reflecting the degree of signal waveform distortion) are obtained through numerical calculation. For example, the mean of the discharge current is μ=15A, the variance is σ²=2.3A², the peak value is I_max=30A, and the waveform factor is K=1.8. To address the frequency domain characteristics, a Fast Fourier Transform (FFT) is performed on the preprocessed signal. Spectral data within the frequency range of 0-5kHz is extracted, and the dominant frequency amplitude (the amplitude corresponding to the frequency with the highest energy in the spectrum) and the spectral centroid (the location reflecting the concentration of spectral energy distribution) are extracted. The formula is as follows: ,in For the i-th frequency point, This corresponds to the amplitude.

[0027] S4. Conduct feature importance assessment: The ReliefF algorithm was used to calculate the anomaly discrimination of each feature. The number of neighborhood samples k=10 and the number of iterations=1000 were set to obtain the weight values ​​of each feature: peak discharge current (0.85), peak electrode vibration (0.82), working fluid flow rate (0.72), discharge voltage variance (0.70), electrode temperature (0.68), working fluid pressure (0.65), discharge-electrode synergy feature (0.75), and electrode-working fluid adaptation feature (0.69). Redundant features with a discrimination index below 0.3 are removed (there are no redundant features in this embodiment), and finally 8 core features are retained to form a time-series feature vector.

[0028] An anomaly detection module, which is communicatively connected to the signal processing module, includes a processing stage identification unit and a multimodal temporal anomaly detection model. It is used to output a status identifier representing the processing state based on the time-domain feature set, frequency-domain feature set, and associated fusion feature set. The status identifier includes normal state, early warning state, and alarm state. The intelligent decision-making module, which is communicatively connected to the anomaly detection module, is configured to call a preset three-level mapping error correction strategy library and generate corresponding control commands based on the state identifier and the corresponding anomaly feature combination. The execution control module is communicatively connected to the intelligent decision-making module and the CNC system of the EDM machine, and is used to execute the control commands and verify the error correction effect.

[0029] In this embodiment, the processing stage identification unit constructs a processing stage classification model based on the C4.5 decision tree algorithm. The training dataset consists of 500 samples from different processing stages (100 samples from the initiation stage, 300 samples from the stabilization stage, and 100 samples from the termination stage). The input feature is the discharge frequency (…). ) and electrode feed rate ( ) Decision tree classification rules: If the discharge frequency fluctuates >5Hz and electrode feed rate >0.1mm / s indicates the start-up phase (unstable discharge and rapid electrode feed in the early stages of processing). If the discharge frequency fluctuates ≤2Hz and electrode feed rate 0.02mm / s≤ ≤0.1mm / s, is considered to be in the stable period (uniform discharge, electrode feed at a constant speed); If the discharge frequency fluctuates by 2Hz < ≤5Hz and electrode feed rate <0.02mm / s indicates the finishing stage (processing is nearing completion, electrode is being fed slowly). The unit achieves a recognition accuracy of 98.5% during the processing stage, with a recognition delay of ≤200ms.

[0030] The multimodal temporal anomaly detection model uses the temporal feature vector composed of the aforementioned temporal feature set, frequency domain feature set, and associated fused feature set as input. It includes a Bi-Gated Recurrent Unit (Bi-GRU) network based on an attention mechanism. The input layer dimension is 8 (number of core features). Each of the forward and backward GRU layers contains two hidden layers, with the number of neurons in the hidden layers adapted to the processing stage (64 in the startup phase and 128 in the stable phase). The attention layer uses a single-hidden-layer feedforward neural network (32 hidden layer neurons). The output layer is a Softmax classifier for three states (normal, warning, and alarm). The initial training dataset consists of 10,000 samples (7,000 normal samples, 2,000 warning samples, and 1,000 alarm samples), divided into training and validation sets in a 7:3 ratio. The optimizer is Adam, with a learning rate of 0.005 in the startup phase, 0.001 in the stable phase, and 0.002 in the closing phase. The batch size is 32, the training epochs are 50, and the cross-entropy loss function is used.

[0031] The bidirectional gated recurrent unit network is used to encode the temporal feature vector and calculate the hidden state at each time step; the output layer of the multimodal temporal anomaly detection model maps the final hidden state to the state identifier. The model update unit is configured as follows: Regularly collect real normal samples (with preset normal labels) confirmed by the sample verification unit and labeled real abnormal samples (with clear abnormal type labels) to construct an incremental training set; based on the incremental training set, trigger the incremental learning process of the model, dynamically adjust the parameters of the bidirectional gated recurrent unit network to adapt to the feature distribution drift caused by aging of processing equipment, fluctuations in operating conditions, etc., and optimize the recognition accuracy of various abnormal patterns.

[0032] The sample verification unit is configured as follows: When the status identifier is in a normal state, multi-dimensional consistency verification is performed, such as feature similarity threshold verification with historically labeled normal samples and cross-validation of multi-sensor data, to confirm that the processing process corresponding to the current time-series feature vector is a "true normal working condition". When the status identifier is in a warning or alarm state, record the time-series feature vector and the corresponding processing scenario information. After manual review or fault tracing verification, label the actual abnormal type, such as slight electrode wear or insufficient working fluid flow. Specifically, the time-domain and frequency-domain features extracted by the signal processing module are combined into a multi-dimensional time-series feature vector, which is input into the multimodal time-series anomaly detection model according to the time sequence; the core feature set is input into the model according to the time sequence (time step 100ms, sequence length 20), the forward GRU layer processes the sequence in time order, and the backward GRU layer processes it in time reverse order to obtain the bidirectional hidden state at each time step; The model uses a bidirectional gated recurrent unit (Bi-GRU) network to encode the sequence. GRU units can effectively capture long-short-term dependencies in time series, while the bidirectional structure can simultaneously utilize past and future contextual information to determine the current state, which is crucial for identifying gradual anomalies (such as slow electrode wear) caused by the accumulation of multiple factors.

[0033] After the Bi-GRU outputs the bidirectional hidden state at each time step, the attention mechanism begins to work. It calculates the importance weight (attention score) for each time step using a small neural network. For example, it might find that the sudden changes in the "electrode vibration peak" and "discharge current spectrum centroid" at a certain moment contribute the most to the current judgment of "warning state". These weights are automatically weighted and summarized from the information of all time steps to generate a context vector that focuses on the most relevant information.

[0034] The final context vector is fed into a fully connected layer and a Softmax layer, which are mapped to the probability distribution of three states: "normal", "warning", and "alarm". The state identifier with the highest probability is the output state identifier.

[0035] Optionally, the computation process of the attention-based bidirectional gated recurrent unit network includes: The forward computation layer processes temporal feature vectors in chronological order to obtain the forward hidden state sequence. ; The backward computation layer processes the temporal feature vectors in reverse chronological order to obtain the backward hidden state sequence. ; Each time step The forward hidden state and the backward hidden state are concatenated to obtain the hidden state containing complete context information. ; Calculate the hidden state at each time step attention weights The hidden states at all time steps are summed in weight according to their attention weights to obtain the final context vector. , , This represents the total number of time steps of the processed temporal feature vectors; The final context vector It is fed into the output layer for classification.

[0036] By adopting the above technical solutions, the attention mechanism can provide the basis for the model to make decisions, that is, to visually show which features at which time points caused the "warning" or "alarm". This provides operators with valuable diagnostic information, making it easier to quickly locate the root cause of the problem (e.g., whether the electrode vibration was abnormal first or the discharge was abnormal first). It realizes the functional upgrade from detection to auxiliary diagnosis, and has a higher degree of intelligence.

[0037] Optionally, the attention weights The calculation methods include: Where j represents the loop counter variable, indicating that each time step from 1 to T is traversed during the summation process. Represents the hidden state at time step j. The attention scoring function is... This is achieved through a feedforward neural network, which includes fully connected layers and nonlinear activation functions. ; The calculation formula for the attention scoring function includes: in, It is a trainable weight matrix. It is a trainable weight vector used to compress the mapped result into a scalar score. It is the hyperbolic tangent activation function, used to introduce nonlinearity. It is a trainable bias vector.

[0038] By adopting the above technical solution, the trainable feedforward neural network can learn the dynamic weight allocation strategy of different signal characteristics at different times during the EDM process through data. For example, the current stability is most important at the beginning of the EDM discharge, while the frequency characteristics of electrode vibration are more discriminative in the middle of the process. This dynamic and adaptive feature importance assessment capability at different stages is one of the core advantages of this solution compared with traditional fixed threshold or static models.

[0039] Optionally, the anomaly detection module includes an output module, and the output module's status identifier output rule is as follows: Output normal status when all features are within the normal range; A warning status is output when at least one feature exceeds the normal range but does not reach the alarm threshold. An alarm status is output when any feature reaches or exceeds the alarm threshold.

[0040] Optionally, it may also include a display alarm module and a data storage module; The alarm display module integrates a touchscreen, displaying waveforms, characteristic values, and status indicators in real time. A yellow indicator flashes during a warning, and a red indicator flashes and a buzzer sounds during an alarm, providing intuitive human-machine interaction for timely operator response. The data storage module uses an SQL database to record all sensor data, feature values, state changes, and error correction actions. Data can be queried and exported by time for subsequent analysis and system optimization. It supports long-term operational evaluation and fault diagnosis.

[0041] This application also discloses a method for detecting and intelligently correcting anomalies in the EDM machining process, which is applied to the EDM machining process anomaly detection and intelligent error correction system described in any of the above embodiments.

[0042] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A fault detection and intelligent error correction system for EDM machining process, characterized in that, include: The sensor module is installed in the discharge circuit, electrode system and working fluid system of the EDM machine to collect multi-source time-series data in real time during the processing. The multi-source time-series data includes discharge voltage, discharge current, electrode vibration signal, electrode temperature signal, working fluid flow rate and working fluid pressure. The signal processing module is communicatively connected to the sensor module and is used to preprocess, feature association model, and feature extraction of the multi-source time-series data to obtain a time-domain feature set, a frequency-domain feature set, and an association-fused feature set. An anomaly detection module, which is communicatively connected to the signal processing module, includes a processing stage identification unit and a multimodal temporal anomaly detection model. It is used to output a status identifier representing the processing state based on the time-domain feature set, frequency-domain feature set, and associated fusion feature set. The status identifier includes normal state, early warning state, and alarm state. The intelligent decision-making module, which is communicatively connected to the anomaly detection module, is configured to call a preset three-level mapping error correction strategy library and generate corresponding control commands based on the state identifier and the corresponding anomaly feature combination. The execution control module is communicatively connected to the intelligent decision-making module and the CNC system of the EDM machine, and is used to execute the control commands and verify and provide feedback on the error correction effect.

2. The EDM machining process anomaly detection and intelligent error correction system according to claim 1, characterized in that, The signal processing module includes a feature correlation analysis unit and a feature importance evaluation unit, specifically configured as follows: Preprocessing unit: Employs a differentiated adaptive filtering algorithm to denoise multi-source time-series data; Feature association analysis unit: Calculates the association degree of multi-source features based on mutual information entropy, establishes an association matrix, and performs weighted fusion of strongly associated features to obtain associated fused features; Feature extraction unit: calculates time-domain features and frequency-domain features. The time-domain features include mean, variance, peak value and waveform factor. The frequency-domain features include the main frequency amplitude and spectral centroid obtained after performing a fast Fourier transform on the preprocessed signal. Feature importance evaluation unit: The ReliefF algorithm is used to calculate the discriminative power of each feature, and redundant features with discriminative power below a specified value are removed, while the core feature set is retained.

3. The EDM machining process anomaly detection and intelligent error correction system according to claim 2, characterized in that, The anomaly detection module includes a multimodal temporal anomaly detection model; The multimodal temporal anomaly detection model takes the temporal feature vector composed of the temporal feature set, frequency feature set, and correlation fusion feature set as input, and includes a bidirectional gated recurrent unit network based on attention mechanism, a sample verification unit, and a model update unit. The bidirectional gated recurrent unit network is used to encode the temporal feature vector and calculate the hidden state at each time step; the output layer of the multimodal temporal anomaly detection model maps the final hidden state to the state identifier. The sample verification unit is configured as follows: When the status identifier is in a normal state, the processing procedure corresponding to the current time series feature vector is confirmed to be "true normal working condition" through multi-dimensional consistency verification. When the status identifier is a warning or alarm status, record the time-series feature vector and the corresponding processing scenario information. After manual review or fault tracing verification, label its true abnormal type. The model update unit is configured as follows: Regularly collect real normal samples confirmed by the sample verification unit and labeled real abnormal samples to construct an incremental training set; based on this incremental training set, trigger the incremental learning process of the model and dynamically adjust the parameters of the bidirectional gated recurrent unit network.

4. The EDM machining process anomaly detection and intelligent error correction system according to claim 3, characterized in that, The computation process of the attention-based bidirectional gated recurrent unit network includes: The forward computation layer processes temporal feature vectors in chronological order to obtain the forward hidden state sequence. ; The backward computation layer processes the temporal feature vectors in reverse chronological order to obtain the backward hidden state sequence. ; Each time step The forward hidden state and the backward hidden state are concatenated to obtain the hidden state containing complete context information. ; Calculate the hidden state at each time step attention weights The hidden states at all time steps are summed in weight according to their attention weights to obtain the final context vector. , , This represents the total number of time steps of the processed temporal feature vectors; The final context vector It is fed into the output layer for classification.

5. The EDM machining process anomaly detection and intelligent error correction system according to claim 4, characterized in that, The attention weight The calculation methods include: Where j represents the loop counter variable, indicating that each time step from 1 to T is traversed during the summation process. Represents the hidden state at time step j. This is the attention scoring function.

6. The EDM machining process anomaly detection and intelligent error correction system according to claim 5, characterized in that, The attention scoring function is implemented using a feedforward neural network, which includes fully connected layers and non-linear activation functions. ; The calculation formula of the attention scoring function include: in, It is a trainable weight matrix. It is a trainable weight vector used to compress the mapped result into a scalar score. It is the hyperbolic tangent activation function, used to introduce nonlinearity. It is a trainable bias vector.

7. The EDM machining process anomaly detection and intelligent error correction system according to claim 3, characterized in that, The anomaly detection module includes an output module, and the output module's status identifier output rule is as follows: Output normal status when all features are within the normal range; A warning status is output when at least one feature exceeds the normal range but does not reach the alarm threshold. An alarm status is output when any feature reaches or exceeds the alarm threshold.

8. The EDM machining process anomaly detection and intelligent error correction system according to claim 1, characterized in that, It also includes an alarm display module and a data storage module; The display alarm module is used to display the processing status, feature data, status identifier and error correction action in real time, and to provide audible and visual alarms in the early warning and alarm states. The data storage module is used to store multi-source time-series data, feature data, historical records of status identifiers, and error correction logs, and supports data backtracking and analysis.

9. A method for abnormal detection and intelligent error correction in EDM machining process, characterized in that, It is applied to the abnormal detection and intelligent error correction system for EDM machining process as described in any one of claims 1-8.

Images