An online monitoring system for grinding machine working process

Abstract

The application relates to the field of grinder online detection technology and discloses a grinder working process online monitoring system, which comprises a data acquisition module, vibration, acoustic emission and environment signals are acquired, and original multi-dimensional signal vectors are generated; a data preprocessing module, the original multi-dimensional signal vectors are preprocessed, and a purified signal set is generated; a working condition inversion module, real-time inversion of current grinding process parameter estimation values; an abnormality detection module, based on the purified signal set, the abnormality score quantifying the abnormality degree of the current working condition is calculated; a self-adaptive adjustment module, according to the process parameters, a dynamic threshold value is generated, and the abnormality score is compared to determine the abnormality state; a decision fusion module, the abnormality state and the environment humidity are fused, and a decision instruction is generated. The application estimates the material hardness of the processed workpiece in real time, dynamically adjusts the abnormality threshold value according to the hardness estimation value, effectively distinguishes the normal signal fluctuation caused by normal working condition switching and equipment failure, and avoids false alarms caused by process changes.

Description

Technical Field

[0001] This application relates to the field of online monitoring technology for grinding machines, specifically an online monitoring system for the working process of a grinding machine. Background Technology

[0002] Grinding is a key process in precision manufacturing, and its processing stability has a decisive impact on the final dimensional accuracy and surface quality of the workpiece. In order to detect and diagnose abnormal conditions such as grinding wheel wear and workpiece burns in a timely manner during the grinding process, a variety of online monitoring systems have been developed in this field.

[0003] In existing grinding machine monitoring systems, a typical technical solution involves installing piezoelectric accelerometers or resonant acoustic emission sensors on key components of the grinding machine to collect high-frequency vibration and acoustic signals during the machining process. After acquiring these raw signals, the system typically uses a data preprocessing unit, such as a fixed-parameter digital bandpass filter, to filter out out-of-band power frequency interference or high-frequency random noise. After obtaining the purified signal, the system extracts its time-domain or frequency-domain characteristics and compares these characteristics with a pre-set, static anomaly threshold. Once the characteristic value exceeds the threshold, the system determines that an anomaly has occurred and triggers an audible and visual alarm or executes a shutdown command.

[0004] However, in actual processing, batch changes of workpieces or variations in material hardness are common in existing detection systems. These normal changes in operating conditions cause natural spikes in grinding signal characteristics. Fixed thresholds are difficult to adapt to these changes, easily misinterpreting normal operating condition changes as equipment malfunctions, thus frequently triggering false alarms. Furthermore, background noise in the workshop environment is dynamically changing, and filters with fixed parameters struggle to match optimal filtering parameters in real time, leading to a reduced signal-to-noise ratio in noisy environments and affecting monitoring reliability. Moreover, existing feature threshold comparisons can only output a yes / no judgment, failing to integrate the severity of the anomaly, the specific type of anomaly, and crucial external environmental information such as workshop humidity. Therefore, this invention provides an online monitoring system for the grinding machine's operating process to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this application is to provide an online monitoring system for the working process of a grinding machine. This system solves the problems of existing grinding machine monitoring systems lacking adaptability and prone to false alarms when facing complex and ever-changing actual working conditions and environmental interference; at the same time, their decision-making logic is simple and unable to integrate multi-dimensional information for context-aware intelligent decision-making.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an online monitoring system for the working process of a grinding machine, comprising:

[0007] The data acquisition module is used to acquire vibration signals, acoustic emission signals, ambient humidity and noise signals during the grinding process in real time, so as to generate raw multidimensional signal vectors;

[0008] The data preprocessing module is used to preprocess the original multidimensional signal vector to generate a cleaned signal set;

[0009] The working condition inversion module is used to invert the estimated values ​​of process parameters for the current grinding process in real time based on the purification signal set.

[0010] An anomaly detection module is used to calculate an anomaly score that quantifies the degree of anomaly in the current operating condition based on the purification signal set.

[0011] The adaptive adjustment module is used to generate a dynamic anomaly threshold based on the process parameter estimates obtained by the operating condition inversion module, and compare the dynamic anomaly threshold with the anomaly score calculated from the anomaly detection module to determine the anomaly state of the current working process.

[0012] The decision fusion module is used to fuse the abnormal state and the environmental humidity signal obtained from the data acquisition module, and generate decision instructions based on a preset decision logic library.

[0013] Preferably, the operating condition inversion module includes:

[0014] The offline training unit is used to extract feature vectors from the set of purification signals with working condition labels, and to train the model using the feature vectors as input and the known material hardness value as a label, through a supervised learning algorithm to obtain a fixed working condition inversion model.

[0015] The online inversion unit is used to call the solidified working condition inversion model, input the extracted feature vector into the working condition inversion model, and calculate the estimated values ​​of process parameters, wherein the estimated values ​​of process parameters are the estimated values ​​of material hardness.

[0016] Preferably, the anomaly detection module includes:

[0017] The feature engineering unit is used to transform the purified signal set generated by the data preprocessing module into a real-time feature vector.

[0018] The health status assessment unit is used to input the real-time feature vector into a pre-trained deep autoencoder model to obtain a reconstruction vector, and to obtain the anomaly score by calculating the reconstruction error between the real-time feature vector and the reconstruction vector.

[0019] Preferably, the deep autoencoder model is composed of an encoder and a decoder connected in series;

[0020] The encoder is used to compress the real-time feature vector into low-dimensional latent space features.

[0021] The decision fusion module is also used to fuse the low-dimensional latent space features to generate decision instructions based on the anomaly types represented by the latent space features.

[0022] Preferably, the adaptive adjustment module includes:

[0023] The parameter adjustment unit is used to generate the dynamic anomaly threshold based on the estimated process parameters obtained by the operating condition inversion module, and to generate filtering parameters through a mapping relationship model based on the environmental noise signal obtained by the data acquisition module and output them to the data preprocessing module.

[0024] The status determination unit is used to compare the anomaly score of the anomaly detection module with the dynamic anomaly threshold to determine the anomaly status of the current working process.

[0025] Preferably, the decision fusion module includes:

[0026] The multimodal input integration unit is used to collect and synchronize the binary abnormal state flag from the adaptive adjustment module, the quantitative abnormal score from the abnormal detection module, the latent space features from the abnormal detection module, and the environmental humidity signal from the data acquisition module to form a unified decision input vector.

[0027] The decision logic core unit is used to analyze the decision input vector using a fuzzy logic-based inference engine and output the decision instruction.

[0028] Preferably, the decision logic core unit performs reasoning based on a predefined fuzzy rule base, wherein the rules used in the fuzzy rule base are:

[0029] If the abnormal state is yes, the abnormal score is high, and the ambient humidity is high, then the output decision is to execute an emergency shutdown.

[0030] Preferably, the data acquisition module includes:

[0031] The key component sensing unit includes a piezoelectric accelerometer for acquiring the vibration signal and a resonant acoustic emission sensor for acquiring the acoustic emission signal.

[0032] The wireless sensor network unit consists of multiple wireless sensor nodes and a sink node. The wireless sensor nodes integrate a capacitive humidity sensor for acquiring the ambient humidity and an electret microphone for acquiring the noise signal. The wireless sensor nodes and the sink node transmit data through a low-power wireless communication protocol that supports adaptive frequency hopping technology and / or mesh network topology.

[0033] Preferably, the data preprocessing module is used to execute differentiated processing strategies, including:

[0034] A digital bandpass filter is applied to the vibration signal, and a wavelet threshold denoising method is applied to the acoustic emission signal.

[0035] The temperature signal and the ambient humidity signal are filtered using a moving average filter or an exponentially weighted average filter.

[0036] The present invention also provides a method for online monitoring of the working process of a grinding machine, comprising the following steps:

[0037] S1. Acquire vibration signals, acoustic emission signals, and environmental humidity and noise signals during the grinding process to generate the original multidimensional signal vector;

[0038] S2. Process the original multidimensional signal vector to generate a purified signal set;

[0039] S3. Based on the purification signal set, the estimated values ​​of process parameters related to the current grinding process are retrieved in real time.

[0040] S4. Based on the purification signal set, calculate the anomaly score that quantifies the degree of anomaly in the current working condition;

[0041] S5. Based on the estimated process parameters, generate a dynamic anomaly threshold, and compare the dynamic anomaly threshold with the anomaly score to determine the anomaly status of the current working process.

[0042] S6. Integrate the abnormal state and the environmental humidity signal, and generate a decision instruction based on a preset decision logic library.

[0043] In summary, this application includes at least one of the following beneficial technical effects:

[0044] 1. This invention sets up a working condition inversion module to estimate the material hardness of the workpiece in real time, and enables the adaptive adjustment module to dynamically adjust the abnormal threshold based on the hardness estimate. This allows the system to effectively distinguish between normal signal fluctuations caused by switching normal working conditions, such as changing to workpieces with different hardness, and real equipment failures, thus avoiding false alarms caused by process changes and improving the accuracy of monitoring.

[0045] 2. This invention achieves closed-loop adaptive response to environmental interference through an adaptive adjustment module. This adaptive adjustment module can dynamically adjust the filtering parameters in the data preprocessing module in real time according to the environmental noise signal obtained from the data acquisition module, so that the system can actively suppress the interference of changing external environmental noise, ensure the high signal-to-noise ratio of the purified signal set, and improve the robustness of the system in complex workshop environments.

[0046] 3. The decision fusion module of this invention can integrate multimodal information, including binary abnormal states, quantified abnormal scores, latent space features representing abnormal types, and environmental humidity signals representing the external environment. Through a fuzzy logic-based inference engine, the system can make situation-aware decisions. For example, when additional risks such as high humidity are detected, more prudent decisions (such as emergency shutdown) will be made, thereby improving the intelligence level and safety of the grinding process. Attached Figure Description

[0047] Figure 1 This is the system architecture diagram of this application;

[0048] Figure 2 This is a schematic diagram of the decision fusion module of this application;

[0049] Figure 3 This is a flowchart of the method steps in this application. Detailed Implementation

[0050] The following is in conjunction with the appendix Figure 1 -Appendix Figure 3 This application will be described in further detail below.

[0051] See attached document Figure 1 , Figure 1 This is an architecture diagram of an online monitoring system for the working process of a grinding machine according to an embodiment of the present invention. The online monitoring system for the working process of a grinding machine provided by the present invention includes:

[0052] The data acquisition module includes sensors for monitoring key components of the grinding machine and a wireless sensor network for monitoring the surrounding environment of the grinding machine. The data acquisition module is used to acquire vibration signals, temperature signals, acoustic emission signals, and environmental humidity and noise signals in real time during the operation of the grinding machine, and generate raw multidimensional signal vectors.

[0053] The data preprocessing module, electrically connected to the data acquisition module, is used to filter or reduce noise in the generated raw multidimensional signal vector to generate a purified signal set. The filtering parameters of the data preprocessing module are dynamically set by the adaptive adjustment module.

[0054] The working condition inversion module is electrically connected to the data preprocessing module. It is equipped with an analysis model for identifying specific working conditions. Based on the clean signal set, it can invert the estimated values ​​of process parameters related to the current grinding process in real time, such as the estimated value of material hardness.

[0055] The anomaly detection module, electrically connected to the data preprocessing module, is used to extract features from the purification signal set and calculate an anomaly score that quantifies the degree of anomaly in the current working condition using a preset health status model, as well as latent space features used to characterize the anomaly type.

[0056] The adaptive adjustment module is electrically connected to the data acquisition module, the operating condition inversion module, and the anomaly detection module, respectively, and is used to dynamically adjust the system's analysis strategy based on data from different sources.

[0057] Specifically, the adaptive adjustment module generates and outputs filtering parameters to the data preprocessing module based on the environmental noise signal obtained from the data acquisition module, generates a dynamic anomaly threshold based on the process parameter estimates obtained from the operating condition inversion module, and compares the threshold with the anomaly score obtained from the anomaly detection module to determine the abnormal state of the current working process.

[0058] The decision fusion module is electrically connected to the adaptive adjustment module, the anomaly detection module, and the data acquisition module. It is used to fuse abnormal states, anomaly scores, hidden space features, and environmental humidity signals, and generate the optimal decision instruction based on a preset decision logic library. This decision instruction is sent to the control unit of the grinding machine to perform operating parameter adjustments or trigger early warnings.

[0059] The data acquisition module includes key components such as a sensing unit and a wireless sensor network unit, which work together to form a multi-source heterogeneous signal acquisition system.

[0060] The key component sensing unit is used to directly monitor the core motion and functional components of the grinding machine. This unit can integrate various types of sensors, including a piezoelectric accelerometer to acquire spindle vibration information. This sensor is magnetically mounted or screwed to the spindle housing near the front bearing, a location that most effectively captures vibration signals caused by bearing wear, grinding wheel imbalance, or grinding abnormalities. To acquire temperature information from key heat-generating components, a K-type thermocouple or a PT100 platinum resistance thermometer can be used. Its probe can be embedded in a pre-drilled temperature measurement hole in the motor winding or attached to the outer wall of the spindle bearing housing using thermally conductive adhesive to monitor its operating temperature rise. To acquire high-frequency stress wave signals during the grinding process, a resonant acoustic emission sensor is also used, fixed to a stationary component such as the workpiece fixture or tailstock near the grinding area to collect acoustic emission signals related to the generation of micro-cracks in the material and the wear state of the grinding wheel.

[0061] The wireless sensor network unit, used to monitor changes in the machining environment surrounding the grinding machine, consists of multiple wireless sensor nodes and a convergence node. To ensure high reliability and real-time performance of data transmission in industrial environments with strong electromagnetic interference, a multi-layered anti-interference and fault-tolerant strategy is adopted.

[0062] The wireless sensor nodes can integrate a capacitive humidity sensor for measuring air humidity and an electret microphone for measuring ambient acoustic noise. These wireless sensor nodes are placed in the area surrounding the grinding machine, with a deployment strategy that prioritizes avoiding strong electromagnetic interference sources such as frequency converters and high-power motors, as well as main personnel activity channels, to ensure the stability of signal transmission.

[0063] To address the challenges of complex wireless channels, wireless sensor network units prioritize low-power wireless communication protocols with high anti-interference capabilities, such as Zigbee or Bluetooth Low Energy Mesh protocols that support mesh network topology and adaptive frequency hopping (AFH) technology. Specifically:

[0064] Adaptive frequency hopping: This technology is a standard feature of Bluetooth Low Energy and Zigbee. The wireless network can monitor the usage and interference levels of each channel in real time and automatically skip congested channels that are interfered with, communicating only on relatively clean channels, thereby avoiding direct conflict with other wireless devices (such as Wi-Fi) or electromagnetic noise in the environment.

[0065] Acknowledgment and retransmission mechanism: At the protocol level, the data sender waits for acknowledgment from the receiver after sending a data packet. If no acknowledgment is received within a preset time, the system determines that the data packet is lost and automatically retransmits it until it is successfully delivered or the maximum number of retransmissions is reached. This effectively solves the problem of data packet loss caused by momentary interference.

[0066] Mesh Networks: By employing a mesh network topology supported by BLE Mesh or Zigbee, each node in the network can not only communicate with the sink node but also act as a relay, forwarding data to other nodes. If the direct link between a node and the sink node is interrupted due to interference, its data can automatically find a new available path through neighboring nodes, enhancing the robustness and coverage of the entire network.

[0067] To address data delays or loss in extreme cases (i.e., after retransmission failure), the receiving end (aggregation node or data preprocessing module) is designed with fault-tolerant logic. If environmental data (such as humidity) fails to arrive on time at a certain moment, the system can employ a short-term predictive filling strategy. For example, it can use the valid value from the previous moment for "zero-order hold" or perform linear interpolation based on historical trends to generate a temporary estimate. This ensures that downstream adaptive adjustment modules and other analysis units do not stop working due to data stream interruption, guaranteeing the continuity of key functions such as dynamic threshold adjustment, although accuracy may decrease slightly in the short term. Simultaneously, the system will mark this filled data and correct it upon the arrival of subsequent real data.

[0068] The signals collected by the key component sensing unit can be connected to the multi-channel synchronous data acquisition card via shielded cables, while the signals collected by the wireless sensor network unit are output by its aggregation node. The system synchronizes and aligns the signal streams from these two sources in the time dimension and integrates them into a unified raw multi-dimensional signal vector containing multiple time series. This vector can be represented by the following formula:

[0069] ;

[0070] in: Indicates the sampling time; This represents the time series of vibration signals collected by an accelerometer. This represents a time series of temperature signals collected by a temperature sensor. This represents the time series of acoustic emission signals collected by the acoustic emission sensor; This represents a time series of ambient humidity signals collected by a wireless sensor network. This represents a time series of environmental noise signals collected by a wireless sensor network.

[0071] The data preprocessing module is used to clean up the raw multidimensional signal vector output by the data acquisition module to suppress noise and interference from irrelevant frequency components, thereby improving the signal-to-noise ratio and accuracy of subsequent analysis.

[0072] In one specific embodiment, the data preprocessing module can execute differentiated processing strategies based on the physical characteristics of different signals. For high-frequency dynamic signals containing rich state information, such as vibration signals and acoustic emission signals, the data preprocessing module employs digital filtering or signal decomposition techniques. For vibration signals, a digital bandpass filter, such as a Butterworth filter or a Chebyshev filter, is implemented, with its passband range set to cover the characteristic fault frequencies of the main components of the grinding machine (such as bearings and gears), to filter out power frequency interference and high-frequency random noise. For acoustic emission signals containing a large number of transient impact components, a wavelet thresholding denoising method can be implemented, which involves performing multi-level wavelet decomposition on the signal, applying soft or hard thresholding functions to the wavelet coefficients of each level, and then reconstructing the signal, thereby effectively removing background noise while preserving the impact characteristics.

[0073] For slowly changing signals such as temperature and ambient humidity, the data preprocessing module can use moving average filtering or exponential weighted average filtering to smooth the signal curve and filter out accidental jumps or glitches introduced by sensors or circuits.

[0074] To the vibration signal Taking digital bandpass filtering as an example, its processing can be represented by the following formula:

[0075] ;

[0076] in, This represents the time sequence of the original input vibration signal; This indicates the selected digital bandpass filter function; This represents the configurable parameters of the filter function, which may specifically include the filter's low cutoff frequency. High cutoff frequency and filter order ; This represents the time series of the purified vibration signal output after filtering.

[0077] In one specific implementation, the aforementioned signal processing parameters are not fixed. The data preprocessing module includes a dynamic parameter interface unit, which receives and parses control commands from the subsequent adaptive adjustment module. When the adaptive adjustment module determines that the preprocessing strategy needs to be adjusted based on changes in the external environment (such as ambient noise), it generates new parameters (e.g., a new filter cutoff frequency or wavelet decomposition level) and updates the parameters of the corresponding signal processing algorithm within the data preprocessing module in real time through the dynamic parameter interface unit. This mechanism enables the preprocessing stage of this system to adapt to environmental changes.

[0078] After the above processing, the data preprocessing module reassembles the purified signals to form a purified signal set. The purification signal set is simultaneously transmitted to the subsequent operating condition inversion module and anomaly detection module, serving as a unified data basis for their analysis.

[0079] The working condition feature inversion module replaces the reliance on manually preset or input process parameters. Through in-depth analysis of real-time sensor signals, it automatically and in real-time inverts the key working condition features of the current grinding process. The module is used to estimate the material hardness of the workpiece being processed in real-time and includes an offline training unit and an online inversion unit.

[0080] The offline training unit's function is to construct and solidify a high-precision working condition inversion model. The specific implementation process is as follows: Multiple sets of standard workpieces with different and known material hardnesses (e.g., Rockwell hardnesses of HRC40, HRC50, and HRC60) are prepared. Under a controlled experimental environment, grinding processes are performed on these workpieces with different hardnesses. The data acquisition module and data preprocessing module collect and generate corresponding cleanup signal sets labeled with working conditions. Features sensitive to changes in material hardness are extracted from these signal sets. For example, statistical features such as kurtosis, margin, and spectral entropy are extracted from cleanup vibration signals, and features such as ring count, root mean square value, and energy are extracted from cleanup acoustic emission signals, forming a high-dimensional feature vector. A supervised learning algorithm, such as Gradient Boosting Decision Tree (GBDT), Support Vector Machine (SVM), or Convolutional Neural Network (CNN), is selected. Using the extracted feature vector as input and the known material hardness value as the label, the model is trained until convergence, ultimately yielding the solidified working condition inversion model.

[0081] The function of the online inversion unit is to invoke the pre-trained operating condition inversion model during actual system operation and perform real-time operating condition parameter estimation. This online inversion unit receives a real-time purified signal set from the data preprocessing module, performs a feature extraction process consistent with the offline training phase, and generates a real-time feature vector. This feature vector is then input into the fixed operating condition inversion model for calculation. This calculation process can be represented by the following formula:

[0082] ;

[0083] in, Indicates the current moment; This represents the feature vector extracted from the purification signal set at the current moment for use in operating condition inversion; This represents the working condition inversion model function solidified by the offline training unit; This represents the internal parameter set of the solidified model; This represents the real-time material hardness estimate output by the working condition inversion module.

[0084] The material hardness estimate output by the working condition characteristic inversion module Instead of being presented directly to the user as a monitoring result, this parameter is transmitted in real time to the subsequent adaptive adjustment module as a key internal state parameter. This design enables the system to self-perceive changes in its processed objects without external intervention, and provides a reliable and automated decision-making basis for dynamically adjusting the sensitivity of anomaly detection, thus forming the foundation for the system's closed-loop adaptive capability.

[0085] The anomaly detection module performs in-depth analysis of the set of purification signals characterizing the current working state of the grinding machine to quantify the degree of deviation of the current working condition from the normal state in real time, and generates auxiliary information to characterize the anomaly type. This anomaly detection module includes a feature engineering unit and a health status assessment unit.

[0086] The feature engineering unit transforms the time-series set of cleaned signals from the data preprocessing module into a structured feature vector containing multi-dimensional information. This unit can calculate and combine various types of features, such as: extracting time-domain features from the cleaned vibration signals, including root mean square (RMS), kurtosis, and margin; and frequency-domain features, including spectral entropy and power spectral centroid. Simultaneously, it extracts parameters such as energy and ring count from the cleaned acoustic emission signals. These features, combined with statistical values ​​from signals such as temperature and humidity, together form a snapshot comprehensively describing the system state at the current moment—the feature vector.

[0087] The health status assessment unit analyzes the real-time feature vectors generated by the feature engineering unit to evaluate their degree of abnormality. This unit is based on a pre-trained health status model, such as a deep autoencoder. This deep autoencoder model consists of an encoder and a decoder connected in series. Its basic principle is that by training on a large amount of clean, normal operating condition data, the model can accurately learn the distribution pattern of normal operating condition data in the feature space and has the ability to reconstruct normal feature vectors with low error. When a feature vector under abnormal operating conditions is input, because the feature vector deviates from the normal pattern learned by the model, the model will be unable to reconstruct it effectively, resulting in reconstruction error.

[0088] The encoder part of this deep autoencoder model consists of multiple stacked fully connected layers, used to progressively compress the high-dimensional input feature vector into low-dimensional latent space features. The decoder part has a structure symmetrical to the encoder, used to decode the latent space features layer by layer, attempting to reconstruct the original feature vector. The training process of this deep autoencoder model is offline, and its goal is to minimize the difference between the input vector and the reconstructed vector, such as the mean squared error.

[0089] During system online operation, the health status assessment unit inputs real-time feature vectors into a pre-trained deep autoencoder model to obtain reconstructed vectors. Subsequently, by calculating the squared Euclidean distance between the two vectors, an anomaly score quantifying the degree of current operational anomaly is obtained. This calculation process can be expressed by the following formula:

[0090] ;

[0091] in, Indicates the current moment; This represents the input feature vector at the current moment; This represents the model's reconstruction result for the current feature vector; This represents the square of the L2 norm of the vector, i.e., the reconstruction error; This indicates the current anomaly score.

[0092] The anomaly detection module ultimately outputs two key pieces of information:

[0093] One is the anomaly score obtained from the above calculation. ;

[0094] Secondly, the low-dimensional latent space features of the encoder output. Anomaly scoring It is then passed to a subsequent adaptive adjustment module for comparison with dynamically generated anomaly thresholds. Latent space features Because it encapsulates the core information of the original high-dimensional features and can reflect the inherent type of the abnormal pattern, it is also transmitted to the subsequent decision fusion module as one of the important bases for making complex decisions.

[0095] The adaptive adjustment module is used to dynamically and in a closed loop self-tune the system's analysis and processing links based on the operating condition information retrieved internally and the environmental information perceived externally. This adaptive adjustment module includes a parameter adjustment unit and a state determination unit.

[0096] The parameter adjustment unit is used to perform two parallel adjustment tasks:

[0097] Based on environmental change-based preprocessing parameter adjustment, this parameter adjustment unit receives environmental noise signals from the data acquisition module and has an internal mapping model that predefines the correspondence between different environmental noise levels and specific filter parameters in the data preprocessing module. For example, this mapping model can be a lookup table or a function that maps the real-time measured root mean square value of environmental noise to the cutoff frequency of the optimal high-pass filter for the acoustic emission signal. When environmental noise increases, the parameter adjustment unit automatically calculates a higher cutoff frequency and sends this new parameter through the dynamic parameter interface reserved in the data preprocessing module, thereby actively suppressing the interference of environmental noise on the acoustic emission signal.

[0098] Based on the dynamic adjustment of the abnormal threshold using working condition inversion, this parameter adjustment unit receives real-time material hardness estimates from the working condition feature inversion module. In grinding, different material hardnesses generate "normal" signals with varying amplitudes; higher hardness typically results in stronger vibration or acoustic emission signals. To avoid false alarms caused by normal working condition changes (e.g., replacing with a harder workpiece), this unit uses a preset adjustment function to dynamically calculate a matching abnormal threshold based on the real-time inverted hardness value. This calculation process can be expressed by the following formula:

[0099] ;

[0100] in, Indicates the current moment; This represents the real-time material hardness estimate output by the working condition inversion module; This represents a monotonically increasing function, such as a linear or exponential function, used to map hardness values ​​to an adjustment amount. This represents a preset sensitivity coefficient used to scale the degree of influence of the adjustment amount; This represents a baseline abnormal threshold set under standard operating conditions; This represents the final generated dynamic anomaly threshold that changes in real time with the operating conditions.

[0101] The status determination unit is used to make a final determination on whether the current operating condition is abnormal after the threshold adjustment is completed. The input of the status determination unit includes the real-time abnormality score from the abnormality detection module. and the dynamic anomaly threshold generated by the parameter adjustment unit. The two input values ​​are compared using a comparator. If an anomaly is detected, a score is awarded. Greater than the dynamic anomaly threshold If the condition is met, the current state is determined to be abnormal, and the abnormal state flag is output as "1"; otherwise, it is determined to be normal, and the abnormal state flag is output as "0".

[0102] The binary anomaly status flag output by the adaptive adjustment module is then transmitted to the subsequent decision fusion module. Through this mechanism, the self-awareness of operating conditions is tightly coupled with the anomaly detection process, enabling the entire system to automatically adapt to changes in process parameters, thus improving the accuracy and robustness of monitoring.

[0103] See attached document Figure 2 The decision fusion module integrates multi-source, heterogeneous internal state information with external environmental information, and generates optimal, context-aware decision commands based on preset decision logic. This module includes a multimodal input integration unit and a decision logic core unit.

[0104] The multimodal input integration unit is used to solve the problem of accurate temporal alignment of heterogeneous information from different modules with different generation frequencies and computational delays, ensuring that a time-consistent and distortion-free fused input vector is provided for subsequent decision-making. This multimodal input integration unit specifically includes:

[0105] Unified timestamp: All data within the monitoring system, from the raw signals generated by the data acquisition module to the intermediate results output by each processing module (such as anomaly scores, latent space features, etc.), is assigned a timestamp originating from a unified high-precision system clock.

[0106] Data Buffering and Alignment: The multimodal input integration unit has an independent, timestamped First-In-First-Out (FIFO) buffer for the input signal stream (binary anomaly state, anomaly score, latent space features, ambient humidity). When a new data point arrives, it is stored in the corresponding buffer along with its timestamp.

[0107] Synchronization strategy based on decision clock: The fusion operation is triggered at a preset, fixed "decision clock period" (e.g., every 100 milliseconds). At each decision moment... Perform the following synchronization process to construct the fused input vector. :

[0108] The calculated data (abnormal states, abnormal scores, latent space features) are generated based on the analysis of signals within a time window; therefore, the update frequency is lower than that of the original high-frequency signals at the decision-making time. The system will extract the timestamp that is closest to and no later than the corresponding buffer. The availability of valid data points ensures that the decision-making process uses the latest and most valid analysis results prior to the current decision-making moment.

[0109] For low-frequency sampled environmental signals (such as ambient humidity) The ambient humidity changes very slowly, and the sampling rate is much lower than the system's decision frequency. To obtain an effective humidity value at each decision moment, this multimodal input integration unit employs a time-series interpolation or hold-behind strategy. For example, linear interpolation can be used, based on the data in the buffer... Calculate using the two most recent humidity sampling points. The estimated humidity value at any given time; or, similarly, the nearest neighbor sampling method can be used to directly use the value of the closest timestamp. The sampled values.

[0110] Latency Compensation: From signal acquisition to feature extraction and anomaly score calculation, the entire processing chain inherently involves computational latency. Through a unified timestamp mechanism, the decision fusion module compensates for latency at the decision-making moment. The integrated data, although based on earlier (e.g.) ,in To handle the delay in the original signal, but since all derived data (anomaly scores, status flags, etc.) originate from the same batch of preprocessed signals, their temporal relationships are relatively consistent. The synchronization mechanism of this unit ensures that the ambient humidity used during fusion is also consistent with the time point of the original signal that triggered the computation. The most suitable one is chosen, thus ensuring the causal consistency of the fused input.

[0111] After time synchronization and alignment processing, the multimodal input integration unit generates a fused input vector that is completely synchronized in time. The fused input vector is represented by the following equation:

[0112] ;

[0113] in, Indicates the first A decision-making moment; Indicates the moment of decision-making Aligned binary exception status flags; Indicates the moment of decision-making Aligned anomaly scores; Indicates the moment of decision-making Aligned latent space feature vectors; Indicates the moment of decision-making Aligned environmental humidity signal time series.

[0114] The core decision logic unit analyzes the integrated input vector and outputs the final control command. This core unit employs a fuzzy logic-based inference engine. First, multiple precise numerical values ​​(such as anomaly scores and humidity values) are fuzzified, converted into fuzzy linguistic variables such as "high," "medium," and "low." The fuzzy inference engine then performs inference based on a predefined fuzzy rule base containing multiple "IF-THEN" rules.

[0115] For example, an exemplary rule could be set as follows: if the "abnormal state" is "yes," the "abnormal score" is "high," and the "ambient humidity" is "low," then the output decision is "reduce feed rate." Another rule embodying the innovative idea of ​​this invention could be set as follows: if the "abnormal state" is "yes," the "abnormal score" is "high," but the "ambient humidity" is "high," then the output decision is "execute emergency shutdown." Comparing these two rules, it can be seen that even when facing the same abnormal equipment state, the core unit of this decision logic can make a more prudent and safer decision by perceiving the potential additional risks (such as electrical safety risks) brought about by the external environment (high humidity). Similarly, the rule base can also include rules based on latent space features. The rules are designed to respond differently to different types of anomalies.

[0116] The fuzzy inference engine defuzzifies the fuzzy decision result obtained from the inference (such as "emergency stop") and converts it into explicit instruction code, i.e., decision instructions, that can be recognized and executed by the programmable logic controller (PLC) or CNC system of the grinding machine. The decision command is sent to the grinding machine's control unit to automatically adjust the grinding machine's operating parameters or trigger an audible and visual warning of the corresponding level.

[0117] See attached document Figure 3 This application will describe in detail the online monitoring system for the grinding machine working process provided in the embodiments, in conjunction with a specific implementation scenario. The implementation scenario is set to include two consecutive stages: the first stage is the normal working condition switching, and the second stage is the occurrence of a real fault under a specific environment.

[0118] Phase 1: The grinding machine first performs fine grinding on a batch of workpieces with a Rockwell hardness of HRC40. The system is in a stable monitoring state. The operator changes the processing task and starts to perform the same fine grinding on a new batch of workpieces with a Rockwell hardness of HRC60. Due to the significant increase in the hardness of the processed material, the grinding force increases, resulting in a significant jump in the amplitude and energy of the vibration signal of the grinding machine spindle and the acoustic emission signal in the grinding area.

[0119] During this process, the data acquisition module captures signal changes in real time and transmits them to the data preprocessing module. After receiving the clean signal set, the anomaly detection module calculates a higher anomaly score because the signal characteristics deviate from the normal baseline under the original HRC40 operating condition. The operating condition feature inversion module, which works in parallel with the anomaly detection module, also outputs a material hardness estimate corresponding to HRC60 after analysis by its internal operating condition inversion model (e.g., a pre-trained GBDT model) based on the same signal change.

[0120] The adaptive adjustment module receives two key pieces of information: a higher anomaly score and a new hardness estimate. Its internal parameter adjustment unit immediately calculates a higher dynamic anomaly threshold, matching the HRC60 operating condition, based on a preset adjustment function. Its status determination unit compares the anomaly score with this new threshold and finds that the score does not exceed the dynamic anomaly threshold. Therefore, the system determines that the signal spike was caused by a normal operating condition switch, and ultimately outputs an anomaly status flag of "0" (normal). The system thus successfully avoids a false alarm and seamlessly adapts to the new processing condition.

[0121] Phase Two: The system maintained stable monitoring under HRC60 operating conditions. After a period of time, due to progressive and severe wear of the grinding wheel, the grinding vibration intensified abnormally, and the signal characteristics changed significantly again. Simultaneously, due to weather conditions, the relative humidity in the workshop rose sharply, which was monitored in real time by the wireless sensor network unit.

[0122] The anomaly score calculated by the anomaly detection module increased sharply, exceeding the dynamic anomaly threshold that had been adjusted for the HRC60 operating condition. Simultaneously, the latent space characteristics output by the anomaly detection module pointed to the "grinding wheel wear" fault type in terms of pattern. Based on this, the adaptive adjustment module determined the current state to be a genuine anomaly and output an anomaly status flag of "1".

[0123] The multimodal input integration unit of the decision fusion module receives the complete fusion input vector, including: an abnormal state flag "1", a high-value abnormal score, latent space features representing "grinding wheel wear", and a high humidity signal from the data acquisition module. Its decision logic core unit (e.g., a fuzzy logic engine) begins reasoning. Based on conventional rules that do not include environmental information, the output might be: "If the 'abnormal state' is 'yes' and the 'fault type' is 'grinding wheel wear' and the 'ambient humidity' is 'high,' then the decision instruction is 'emergency stop'." The logical basis of this rule is that a high humidity environment combined with equipment failure may cause secondary risks such as electrical safety or coolant performance failure, requiring the highest level of safety response. Therefore, the module ultimately outputs an "emergency stop" decision instruction to the grinding machine PLC, executing a shutdown operation and illuminating the highest-level alarm light.

Claims

1. An online monitoring system for the working process of a grinding machine, characterized in that, include: The data acquisition module is used to acquire vibration signals, acoustic emission signals, ambient humidity and noise signals during the grinding process in real time, so as to generate raw multidimensional signal vectors; The data preprocessing module is used to preprocess the original multidimensional signal vector to generate a cleaned signal set; The working condition inversion module is used to invert the estimated values ​​of the current grinding process parameters in real time based on the purification signal set; An anomaly detection module is used to calculate an anomaly score that quantifies the degree of anomaly in the current operating condition based on the purification signal set. The adaptive adjustment module is used to generate a dynamic anomaly threshold based on the process parameter estimates obtained by the operating condition inversion module, and compare the dynamic anomaly threshold with the anomaly score calculated from the anomaly detection module to determine the anomaly state of the current working process. The decision fusion module is used to fuse the abnormal state and the environmental humidity signal obtained from the data acquisition module, and generate decision instructions based on a preset decision logic library; The operating condition inversion module includes: The offline training unit is used to extract feature vectors from the set of purification signals with working condition labels, and to train the model using the feature vectors as input and the known material hardness value as a label, through a supervised learning algorithm to obtain a fixed working condition inversion model. The online inversion unit is used to call the solidified working condition inversion model, input the extracted feature vector into the working condition inversion model, and calculate the estimated values ​​of process parameters, wherein the estimated values ​​of process parameters are estimated values ​​of material hardness. The decision fusion module includes: The multimodal input integration unit is used to collect and synchronize the binary abnormal state flag from the adaptive adjustment module, the quantitative abnormal score from the abnormal detection module, the latent space features from the abnormal detection module, and the environmental humidity signal from the data acquisition module to form a unified decision input vector. The decision logic core unit is used to analyze the decision input vector using a fuzzy logic-based inference engine and output the decision instruction.

2. The online monitoring system for the working process of a grinding machine according to claim 1, characterized in that, The anomaly detection module includes: The feature engineering unit is used to transform the purified signal set generated by the data preprocessing module into a real-time feature vector. The health status assessment unit is used to input the real-time feature vector into a pre-trained deep autoencoder model to obtain a reconstruction vector, and to obtain the anomaly score by calculating the reconstruction error between the real-time feature vector and the reconstruction vector.

3. The online monitoring system for the working process of a grinding machine according to claim 2, characterized in that, The deep autoencoder model consists of an encoder and a decoder connected in series. The encoder is used to compress the real-time feature vector into low-dimensional latent space features. The decision fusion module is also used to fuse the low-dimensional latent space features to generate decision instructions based on the anomaly types represented by the latent space features.

4. The online monitoring system for the working process of a grinding machine according to claim 1, characterized in that, The adaptive adjustment module includes: The parameter adjustment unit is used to generate the dynamic anomaly threshold based on the estimated process parameters obtained by the operating condition inversion module, and to generate filtering parameters through a mapping relationship model based on the environmental noise signal obtained by the data acquisition module and output them to the data preprocessing module. The status determination unit is used to compare the anomaly score of the anomaly detection module with the dynamic anomaly threshold to determine the anomaly status of the current working process.

5. The online monitoring system for the working process of a grinding machine according to claim 1, characterized in that, The core unit of the decision logic performs reasoning based on a predefined fuzzy rule base, and the rules used in the fuzzy rule base are as follows: If the abnormal state is yes, the abnormal score is high, and the ambient humidity is high, then the output decision is to execute an emergency shutdown.

6. The online monitoring system for the working process of a grinding machine according to claim 1, characterized in that, The data acquisition module includes: The key component, the sensing unit, includes a piezoelectric accelerometer for acquiring the vibration signal and a resonant acoustic emission sensor for acquiring the acoustic emission signal. The wireless sensor network unit consists of multiple wireless sensor nodes and a sink node. The wireless sensor nodes integrate a capacitive humidity sensor for acquiring the ambient humidity and an electret microphone for acquiring the noise signal. The wireless sensor nodes and the sink node transmit data through a low-power wireless communication protocol that supports adaptive frequency hopping technology and / or mesh network topology.

7. The online monitoring system for the working process of a grinding machine according to claim 1, characterized in that, The data preprocessing module is used to execute differentiated processing strategies, including: A digital bandpass filter is applied to the vibration signal, and a wavelet threshold denoising method is applied to the acoustic emission signal. The ambient humidity signal is filtered using a moving average filter or an exponentially weighted average filter.

8. A method for online monitoring of the working process of a grinding machine, applied to an online monitoring system for the working process of a grinding machine as described in any one of claims 1-7, characterized in that, Includes the following steps: S1. Acquire vibration signals, acoustic emission signals, and environmental humidity and noise signals during the grinding process to generate the original multidimensional signal vector; S2. Process the original multidimensional signal vector to generate a purified signal set; S3. Based on the purification signal set, the estimated values ​​of process parameters related to the current grinding process are retrieved in real time. S4. Based on the purification signal set, calculate the anomaly score that quantifies the degree of anomaly in the current working condition; S5. Based on the estimated process parameters, generate a dynamic anomaly threshold, and compare the dynamic anomaly threshold with the anomaly score to determine the anomaly status of the current working process. S6. Integrate the abnormal state and the environmental humidity signal, and generate a decision instruction based on a preset decision logic library.

Images