High-sensitivity broken tool detection system and broken tool identification method

Abstract

The application discloses a high-sensitivity broken tool detection system and a broken tool identification method, and the system comprises a multi-source synchronous acquisition module, a multi-stage coupling preprocessing module, a segmented fast-slow dual-mode adaptive baseline self-learning module, an identification module, a multi-stage hierarchical linkage alarm module and a machine tool process interactive synchronization module.The high-sensitivity broken tool detection system and the broken tool identification method disclosed by the application realize high-sensitivity identification of micro tool micro-chipping through the organic combination of multi-source non-invasive sensing synchronous acquisition, multi-stage coupling preprocessing, segmented fast-slow dual-mode adaptive baseline, unsupervised lightweight intelligent identification, multi-source cross verification and five-stage hierarchical linkage, significantly reduce the false alarm and missed alarm rates, can be quickly landed without fault sample labeling, can be directly deployed on edge hardware, and is suitable for unmanned flexible numerical control machining scenes.

Description

Technical Field

[0001] This invention belongs to the field of online monitoring technology for CNC machine tool cutting tools, specifically relating to a high-sensitivity broken tool detection system and a broken tool identification method. Background Technology

[0002] In CNC precision machining scenarios, small-diameter, high-length-to-diameter ratio cutting tools have weak rigidity and are highly susceptible to failures such as tip chipping, edge breakage, and overall fracture due to factors such as cutting force fluctuations, uneven material hardness, and improper process parameter matching. In unattended production modes, if tool breakage failures are not identified in a timely manner, it can lead to batch scrapping of workpieces, impact damage to fixtures and spindles, and significantly increase manufacturing and equipment maintenance costs.

[0003] Existing broken blade detection technologies generally have the following shortcomings: First, sensing solutions have detection blind spots. Single vibration monitoring is difficult to accurately capture the faint characteristics of instantaneous fracture, and single acoustic emission monitoring cannot identify progressive tool wear; some multi-sensor solutions rely on expensive sensors such as optical detection and cutting force detection, which require the tool to be moved to a designated detection position to complete the measurement, occupying the machining cycle and making it impossible to achieve real-time monitoring of the entire cutting process, and the deployment and maintenance costs are high.

[0004] Second, the baseline threshold adaptability is poor. Existing solutions mostly use fixed thresholds or simple dynamic thresholds, which cannot adapt to baseline offsets caused by long-term progressive tool wear, ambient temperature drift, workpiece material fluctuations, and switching of cutting conditions. After changing production or tools, technicians need to repeatedly adjust parameters, making it difficult to adapt to flexible production lines with multiple varieties and small batches.

[0005] Third, intelligent identification faces high barriers to implementation. Most machine learning-based detection solutions adopt supervised learning models, relying on a large number of labeled fault samples; however, samples of micro-chipping and broken tools in industrial settings are scarce, and the cost of collection and labeling is extremely high, making it difficult to quickly deploy models; at the same time, most deep learning models require a lot of computing power and rely on host computers to run, making them impossible to deploy directly at the edge of machine tools.

[0006] Fourth, the false alarm and false negative rates are too high. Existing multi-sensor fusion solutions mostly adopt simple weighted fusion strategies. Noise generated by cutting fluid impact, chip collision, and electromagnetic interference in a single channel can easily trigger false alarms, causing frequent unplanned machine tool shutdowns and affecting production cycle time.

[0007] Fifth, the damage grading mechanism is crude. Existing solutions mostly only divide the damage into two or three levels: normal, wear, and fracture. They do not set up an independent early warning level for micro-chipping of the tool tip. This either leads to excessive downtime and reduced production efficiency, or it only responds when the tool is completely broken. It is impossible to achieve early intervention and it is difficult to balance production efficiency and machining quality and safety. Summary of the Invention

[0008] The main objective of this invention is to provide a high-sensitivity broken tool detection system and broken tool identification method. Through the organic combination of multi-source non-invasive synchronous acquisition, multi-stage coupled preprocessing, segmented fast and slow dual-mode adaptive baseline, unsupervised lightweight intelligent identification, multi-source cross-verification, and five-level hierarchical linkage, it achieves high-sensitivity identification of micro-chipping of small tools, significantly reduces false alarm and false negative rates, can be quickly deployed without fault sample labeling, can be directly deployed on edge hardware, and is suitable for unmanned flexible CNC machining scenarios.

[0009] To achieve the above objectives, this invention provides a high-sensitivity broken tool detection system, comprising a multi-source synchronous acquisition module, a multi-stage coupling preprocessing module, a segmented fast / slow dual-mode adaptive baseline self-learning module, an identification module, a multi-level hierarchical linkage alarm module, and a machine tool process interaction synchronization module, wherein: The multi-source synchronous acquisition module includes a three-axis MEMS vibration sensing unit, a wideband acoustic emission (AE) sensing unit, a CNC system power and torque acquisition unit, and a high-precision timestamp alignment unit, which are used to synchronously acquire vibration, acoustic emission, and power and torque signals during the cutting process and unify the timing reference of the three signals. The multi-stage coupling preprocessing module is used to perform adaptive bandpass filtering, dynamic gain adjustment, power frequency electromagnetic interference filtering, and cutting interval screening on the acquired multiple raw signals, eliminating invalid noise and transient interference, and outputting standardized effective characteristic signals for the steady-state cutting stage. The segmented fast and slow dual-mode adaptive baseline self-learning module is used to identify the steady-state cutting range. It adopts a dual-mode mechanism of rapid construction of new tools and slow iteration of continuous machining to generate and update multi-dimensional benchmark feature vectors. At the same time, it freezes the baseline when the working conditions change suddenly and combines spindle temperature compensation to eliminate the baseline offset caused by changes in working conditions and environment. The recognition module is used to extract the multi-dimensional time-frequency features of the three signals. After the reconstruction error is calculated by a lightweight convolutional autoencoder trained only with normal cutting samples, the cross-validation rule of at least two signals synchronously exceeding the limit is executed, and the tool damage level judgment result is output. The multi-level hierarchical linkage alarm module is used to pre-configure threshold ranges corresponding to multi-level tool damage. After confirming the fault through multiple consecutive frames of anti-shake, it performs linkage control of data retention, audible and visual warning, and machine tool feed pause / spindle emergency stop according to the damage level. The machine tool process interaction and synchronization module is used to store the exclusive baseline templates corresponding to various tools, input tool and cutting process parameters, support manual baseline reset when changing tools and materials, and export fault detection log data, so as to realize the synchronous adaptation between the testing instrument and the machine tool process.

[0010] As a further preferred technical solution to the above technical solution, for the multi-source synchronous acquisition module: The triaxial MEMS vibration sensing unit is fixed to the spindle housing of the CNC machine tool and is used to collect the X / Y / Z three-axis vibration timing signals during the cutting process. The wideband acoustic emission (AE) sensing unit is fitted to the front end face of the spindle and is used to capture high-frequency stress wave signals generated by the propagation of microcracks in the tool, tool tip breakage, and instantaneous fracture. The power and torque acquisition unit of the CNC system is connected to the CNC system of the machine tool via an industrial bus to synchronously read the spindle output power, real-time torque, spindle speed, and feed rate process data. The high-precision timestamp alignment unit adds a unified timing reference to the three heterogeneous signals, eliminating feature misalignment caused by the sampling time difference of multiple channels.

[0011] As a further preferred embodiment of the above technical solution, the multi-stage coupling preprocessing module includes an adaptive segmented bandpass filter unit, a dynamic programmable gain unit, an electromagnetic noise notch filter unit, and a cutting interval screening unit, wherein: The adaptive segmented bandpass filter unit dynamically matches the filter passband corresponding to the tooth passing frequency based on the current spindle speed and the number of tool teeth, separating the cutting characteristic frequency band from the machine tool body motion and cutting fluid impact interference frequency band. The dynamic programmable gain unit automatically adjusts the amplification factor according to the real-time signal-to-noise ratio of the signal, increasing the gain for weak breakdown characteristic signals and decreasing the gain for strong cutting steady-state signals. The electromagnetic noise notch filter unit is used to filter out the power frequency electromagnetic interference generated by the machine tool frequency converter. The cutting interval screening unit identifies and marks the transient intervals of empty feed, entry, and exit based on the linkage between signal fluctuation variance and process parameters, and retains only the effective steady-state cutting data, which are then sent to the segmented fast and slow dual-mode adaptive baseline self-learning module and the identification module respectively.

[0012] As a further preferred embodiment of the above technical solution, the segmented fast and slow dual-mode adaptive baseline self-learning module is connected to the multi-stage coupled preprocessing module, and incorporates a cutting interval identification sub-unit, an initial baseline rapid construction sub-unit, a sliding window slow iteration sub-unit, a working condition change baseline freezing sub-unit, and a temperature drift compensation sub-unit, wherein: The cutting interval identification subunit divides machining intervals into four categories—no-load, entry, steady-state cutting, and exit—based on signal fluctuation characteristics, selecting only steady-state cutting data for baseline updates. The initial baseline rapid construction subunit, for new tools or production changeover conditions, rapidly generates a multi-dimensional reference feature vector based on the statistical feature distribution of steady-state cutting signals within a preset number of frames. The sliding window slow iteration subunit uses a fixed-length sliding window in continuous steady-state machining, updating the reference feature vector frame by frame with a low weighting to adapt to the gradual normal wear of the tool. The working condition abrupt change baseline freezing subunit pauses baseline updates when it detects that the spindle speed or feed rate fluctuations exceed a preset range, resuming iteration only after the process parameters have stabilized for a preset duration, avoiding abnormal data contaminating the reference template. The temperature drift compensation subunit collects the spindle housing temperature in real time and adjusts the reference feature amplitude through a temperature drift correction model to eliminate the overall baseline offset caused by equipment temperature rise.

[0013] As a further preferred embodiment of the above technical solution, the baseline iteration control logic of the segmented fast and slow dual-mode adaptive baseline self-learning module is specifically as follows: The cutting interval identification subunit calculates the signal variance within the sliding window in real time. If the variance is lower than the no-load threshold, it is determined to be an empty stroke. The variance stable interval is marked as steady-state cutting. When a new tool is loaded for the first time or the workpiece material is changed, the fast learning mode is activated to collect the statistical mean and fluctuation range of multiple frames of steady-state cutting data and generate an initial multi-dimensional benchmark feature vector. When the same tool is continuously machining in a steady state, it switches to slow iteration mode. For each new frame of valid signal, the baseline feature vector is updated only with low weight, and the weight of old samples decays synchronously. When the machine tool process interaction synchronization module reads that the spindle speed and feed rate fluctuations exceed the preset range, the working condition sudden change baseline freezing subunit locks the reference template, pauses iteration until the process stabilizes for a preset duration, and then unlocks. The temperature drift compensation subunit reads the spindle temperature sampling value in real time, and corrects the reference characteristic amplitude once for each preset temperature change, eliminating the baseline offset caused by temperature drift.

[0014] As a further preferred embodiment of the above technical solution, the recognition module is connected to a multi-stage coupled preprocessing module and a segmented fast-slow dual-mode adaptive baseline self-learning module, and includes a multi-dimensional time-frequency joint feature extraction unit, a lightweight convolutional autoencoder model, and a three-way reconstruction error joint verification unit, wherein: The multi-dimensional time-frequency joint feature extraction unit extracts time-domain statistical features, frequency band energy features and time-frequency transient features from the three preprocessed signals, and splices them together to form three independent multi-dimensional feature vectors. The lightweight convolutional autoencoder model completes unsupervised training using only normal steady-state cutting samples, without the need to label broken or chipped tool fault samples. It adopts a symmetrical lightweight convolutional structure of encoder and decoder, and characterizes the degree to which the tool state deviates from normal cutting by calculating the reconstruction error between the real-time feature vector and the baseline feature vector. The three-way reconstruction error joint verification unit executes a cross-judgment rule: a single signal reconstruction error exceeding the limit is not judged as a tool fault; tool damage is only confirmed when at least two of the three signals of vibration, acoustic emission, and power synchronously exceed the corresponding level of reconstruction error threshold.

[0015] As a further preferred embodiment of the above technical solution, the multi-level hierarchical linkage alarm module is connected to the identification module and pre-configured with five levels of tool damage corresponding to reconstruction error threshold ranges, which sequentially correspond to Level 1 normal cutting, Level 2 slight uniform wear, Level 3 micro-chipping warning, Level 4 severe chipping, and Level 5 complete tool breakage, wherein: The module has built-in continuous multi-frame anti-shake judgment logic. It is necessary to determine that the fault is valid only if multiple consecutive frames are judged to be of the same damage level, and to filter out instantaneous random interference. It also performs differentiated linkage according to the damage level: Level 2 slight wear only stores data locally, Level 3 micro-chipping triggers local audible and visual warning and retains the original waveform, and Level 4 and Level 5 damage immediately sends a pause or emergency stop command to the machine tool control system.

[0016] As a further preferred embodiment of the above technical solution, the machine tool process interaction and synchronization module is connected to the other modules respectively, and has a built-in tool process database, parameter input unit, baseline manual reset unit, and fault log export unit, wherein: The tooling process database stores exclusive baseline templates for tools of different specifications and materials; The parameter input unit is used to input tool specifications, number of teeth, workpiece material, and cutting process parameters; The baseline manual reset unit supports clearing historical baselines and restarting the quick learning process with one click after tool or material change; The fault log export unit supports exporting the original signals, reconstructed error curves, and corresponding process parameters during abnormal periods for process review and optimization.

[0017] This invention also provides a high-sensitivity method for identifying broken blades, comprising the following steps: Step S1, Process initialization configuration: Input tool parameters, workpiece material and cutting process parameters through the machine tool process interaction synchronization module, bind the tool and dedicated baseline storage space, and complete sensor zero-point calibration and machine tool communication docking. Step S2, Multi-source high-precision synchronous acquisition: After the machine tool starts cutting, the three-axis vibration, acoustic emission and power torque signals are acquired synchronously, the timing is aligned by a unified timestamp, and the raw sampling data is output. Step S3, Multi-stage Coupling Preprocessing Noise Reduction: Adaptively switch the filter frequency band according to the current speed and tool parameters, dynamically adjust the signal gain, filter out power frequency electromagnetic interference, identify and shield the transient intervals of no-load feed and cut-in / cut-out, and only output the effective signal of steady-state cutting. Step S4: Segmented fast and slow dual-mode baseline self-learning: Identify steady-state cutting ranges by analyzing signal fluctuation characteristics; When starting a new tool or the first piece of production change, a rapid learning mode is activated to generate an initial multi-dimensional benchmark feature vector based on multi-frame steady-state data. The continuous steady-state processing is switched to a slow iteration mode, and the baseline is updated frame by frame with low weight through a fixed-length sliding window. When a significant fluctuation in process parameters is detected, the baseline is frozen, and iteration resumes after stabilization. The reference amplitude is corrected in real time based on the spindle temperature to compensate for temperature drift. Step S5: Multi-dimensional time-frequency joint feature extraction: Extract time-domain statistical features, frequency band energy features and time-frequency transient features from the three steady-state effective signals respectively, and splice them to form three independent real-time feature vectors; Step S6, Multi-source cross-validation determination: Input the three real-time feature vectors into the lightweight convolutional autoencoder respectively, calculate the reconstruction error of each relative to the baseline feature vector; execute the cross-validation rule for at least two synchronous over-limits, match the five-level damage level with the error threshold range; and output the final determination result after confirmation by multiple consecutive frames of image stabilization. Step S7, Five-level hierarchical linkage processing: Level 1 Normal Cutting: Continuous monitoring, no additional output; Level 2 Minor Wear: Locally archived data, does not trigger alarms or shutdown; Level 3 micro-chipping: Triggers audible and visual warnings, keeps the machine tool processing, and synchronously stores the original data of the entire channel during the fault period; Level 4 severe chipping, Level 5 complete tool breakage: Immediately issue feed pause and spindle brake commands to the machine tool, continuously sound and light alarms, and lock the machining process; Step S8, Fault Data Archiving: The judgment results of all damage events, original waveforms, reconstructed error curves, and corresponding process parameters are automatically stored in local storage, and can be exported for tool life analysis and cutting process optimization.

[0018] The beneficial effects of this invention are as follows: 1. Significantly improved detection sensitivity. Three types of sensor signals complement each other to cover all types of tool failures, multi-layer preprocessing cleans up effective features, and combined with the weak feature capture capability of unsupervised intelligent recognition, it can stably identify micron-level tool tip micro-chipping damage, which is far superior to traditional threshold detection schemes.

[0019] 2. Significantly reduced false alarm and false negative rates. The multi-source signal cross-checking mechanism, combined with multi-frame anti-jitter logic and a baseline freezing strategy for sudden changes in operating conditions, can effectively isolate false triggers caused by cutting fluid impact, chip collision, electromagnetic interference, and operating condition switching, ensuring long-term stable operation.

[0020] 3. Strong adaptability. The segmented fast and slow dual-mode baseline, combined with temperature drift compensation, automatically adapts to progressive tool wear, changes in ambient temperature, and switching of cutting conditions. Tool changes and production changes do not require manual threshold adjustment, making it suitable for flexible, multi-variety production scenarios.

[0021] 4. Convenient and efficient deployment. It adopts non-invasive sensor installation, which does not require modification of machine tools; the unsupervised model does not require the collection and labeling of fault samples; the lightweight network can run directly on edge industrial control hardware without the need for additional computing power equipment, resulting in a short deployment cycle and low cost.

[0022] 5. Scientifically sound and reasonable graded protection. The five-level damage stratification system not only enables early intervention through micro-collapse warning to avoid serious tool breakage accidents, but also avoids excessive downtime affecting production efficiency, perfectly adapting to unmanned, lights-out production modes.

[0023] 6. Data is traceable and optimizable. All fault data is stored synchronously with process parameters, which can be used to analyze the causes of tool failure and optimize cutting process parameters, forming a positive closed loop of monitoring and optimization. Attached Figure Description

[0024] Figure 1 This is a system schematic diagram of the present invention. Figure 2 This is a flowchart illustrating the present invention. Detailed Implementation

[0025] The following description is intended to disclose the present invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art. The basic principles of the invention defined in the following description can be applied to other embodiments, modifications, improvements, equivalents, and other technical solutions that do not depart from the spirit and scope of the invention.

[0026] In the preferred embodiments of the present invention, those skilled in the art should note that the cutting tools and the like involved in the present invention can be considered as prior art.

[0027] Preferred embodiment.

[0028] like Figure 1As shown, this invention discloses a high-sensitivity broken tool detection system, including a multi-source synchronous acquisition module, a multi-stage coupled preprocessing module, a segmented fast and slow dual-mode adaptive baseline self-learning module, a (multi-source cross-validation lightweight convolutional autoencoder) recognition module, a multi-level hierarchical linkage alarm module, and a machine tool process interaction synchronization module, wherein: The multi-source synchronous acquisition module includes a three-axis MEMS vibration sensing unit, a wideband acoustic emission (AE) sensing unit, a CNC system power and torque acquisition unit, and a high-precision timestamp alignment unit, which are used to synchronously acquire vibration, acoustic emission, and power and torque signals during the cutting process and unify the timing reference of the three signals. The three-axis MEMS vibration sensing unit is fixed to the spindle housing of the CNC machine tool and is used to collect the vibration timing signals of the X / Y / Z axes during the cutting process. The three-axis MEMS vibration sensing unit adopts a magnetic non-intrusive installation, and the broadband acoustic emission (AE) sensing unit is installed by coupling agent, neither of which requires modification of the machine tool body structure. The high-precision timestamp alignment unit adopts a precise clock synchronization protocol to ensure that the timing deviation of the three signals is controlled within the microsecond level.

[0029] The wideband acoustic emission (AE) sensing unit is fitted to the front end face of the spindle and is used to capture high-frequency stress wave signals generated by the propagation of microcracks in the tool, tool tip breakage, and instantaneous fracture. The power and torque acquisition unit of the CNC system is connected to the CNC system of the machine tool via an industrial bus to synchronously read the spindle output power, real-time torque, spindle speed, and feed rate process data. The high-precision timestamp alignment unit adds a unified timing reference to the three heterogeneous signals, eliminating feature misalignment caused by the sampling time difference of multiple channels.

[0030] The multi-stage coupling preprocessing module is electrically connected to the multi-source synchronous acquisition module. It is used to perform adaptive bandpass filtering, dynamic gain adjustment, power frequency electromagnetic interference filtering and cutting interval screening on the acquired multi-channel raw signals, eliminate invalid noise and transient interference, and output standardized effective characteristic signals of the steady-state cutting stage. The multi-stage coupling preprocessing module includes an adaptive segmented bandpass filter unit, a dynamically controlled gain unit, an electromagnetic noise notch filter unit, and a cutting interval screening unit, wherein: The adaptive segmented bandpass filter unit dynamically matches the filter passband corresponding to the tooth passing frequency based on the current spindle speed and the number of tool teeth, separating the cutting characteristic frequency band from the machine tool body motion and cutting fluid impact interference frequency band. The dynamic programmable gain unit automatically adjusts the amplification factor according to the real-time signal-to-noise ratio of the signal, increasing the gain for weak breakdown characteristic signals and decreasing the gain for strong cutting steady-state signals. The electromagnetic noise notch filter unit is used to filter out the power frequency electromagnetic interference generated by the machine tool frequency converter. The cutting interval screening unit identifies and marks the transient intervals of empty feed, entry, and exit based on the linkage between signal fluctuation variance and process parameters, and retains only the effective steady-state cutting data, which are then sent to the segmented fast and slow dual-mode adaptive baseline self-learning module and the identification module respectively.

[0031] The segmented fast and slow dual-mode adaptive baseline self-learning module is used to identify the steady-state cutting range. It adopts a dual-mode mechanism of rapid construction of new tools and slow iteration of continuous machining to generate and update multi-dimensional benchmark feature vectors. At the same time, it freezes the baseline when the working conditions change suddenly and combines spindle temperature compensation to eliminate the baseline offset caused by changes in working conditions and environment. The segmented fast and slow dual-mode adaptive baseline self-learning module is connected to the multi-stage coupled preprocessing module, and includes a cutting interval identification sub-unit, an initial baseline fast construction sub-unit, a sliding window slow iteration sub-unit, a working condition change baseline freezing sub-unit, and a temperature drift compensation sub-unit, wherein: The cutting interval identification subunit divides machining intervals into four categories—no-load, entry, steady-state cutting, and exit—based on signal fluctuation characteristics, selecting only steady-state cutting data for baseline updates. The initial baseline rapid construction subunit, for new tools or production changeover conditions, rapidly generates a multi-dimensional reference feature vector based on the statistical feature distribution of steady-state cutting signals within a preset number of frames. The sliding window slow iteration subunit uses a fixed-length sliding window in continuous steady-state machining, updating the reference feature vector frame by frame with a low weighting to adapt to the gradual normal wear of the tool. The working condition abrupt change baseline freezing subunit pauses baseline updates when it detects that the spindle speed or feed rate fluctuations exceed a preset range, resuming iteration only after the process parameters have stabilized for a preset duration, avoiding abnormal data contaminating the reference template. The temperature drift compensation subunit collects the spindle housing temperature in real time and adjusts the reference feature amplitude through a temperature drift correction model to eliminate the overall baseline offset caused by equipment temperature rise.

[0032] The baseline iteration control logic of the segmented fast and slow dual-mode adaptive baseline self-learning module is as follows: The cutting interval identification subunit calculates the signal variance within the sliding window in real time. If the variance is lower than the no-load threshold, it is determined to be an empty stroke. The variance stable interval is marked as steady-state cutting. When a new tool is loaded for the first time or the workpiece material is changed, the fast learning mode is activated to collect the statistical mean and fluctuation range of multiple frames of steady-state cutting data and generate an initial multi-dimensional benchmark feature vector. When the same tool is continuously machining in a steady state, it switches to slow iteration mode. For each new frame of valid signal, the baseline feature vector is updated only with low weight, and the weight of old samples decays synchronously. When the machine tool process interaction synchronization module reads that the spindle speed and feed rate fluctuations exceed the preset range, the working condition sudden change baseline freezing subunit locks the reference template, pauses iteration until the process stabilizes for a preset duration, and then unlocks. The temperature drift compensation subunit reads the spindle temperature sampling value in real time, and corrects the reference characteristic amplitude once for each preset temperature change, eliminating the baseline offset caused by temperature drift.

[0033] The identification module is connected to the multi-stage coupled preprocessing module and the segmented fast and slow dual-mode adaptive baseline self-learning module respectively. It is used to extract the multi-dimensional time-frequency features of the three signals. After the reconstruction error is calculated by the lightweight convolutional autoencoder trained only with normal cutting samples, the cross-validation rule of at least two signals synchronously exceeding the limit is executed, and the tool damage level judgment result is output. The recognition module is connected to a multi-stage coupled preprocessing module and a segmented fast-slow dual-mode adaptive baseline self-learning module, and includes a multi-dimensional time-frequency joint feature extraction unit, a lightweight convolutional autoencoder model, and a three-way reconstruction error joint verification unit, wherein: The multi-dimensional time-frequency joint feature extraction unit extracts time-domain statistical features, frequency band energy features, and time-frequency transient features from the three preprocessed signals, and splices them to form three independent multi-dimensional feature vectors. The time-domain statistical features extracted by the multi-dimensional time-frequency joint feature extraction unit include mean, variance, kurtosis, waveform factor, and peak factor. The frequency band energy features are extracted by Fourier transform to obtain the cumulative energy of the frequency band at the tooth-passing frequency and its harmonics. The time-frequency transient features are extracted by wavelet packet decomposition to obtain the high-frequency sub-band energy, which is used to capture instantaneous collapse impact.

[0034] The lightweight convolutional autoencoder model uses only normal steady-state cutting samples for unsupervised training, eliminating the need to label broken or chipped tool fault samples. It employs a symmetrical lightweight convolutional structure for both encoder and decoder, allowing deployment on embedded edge hardware. The model characterizes the degree to which the tool state deviates from normal cutting by calculating the reconstruction error between the real-time feature vector and the baseline feature vector. The encoder of the lightweight convolutional autoencoder model consists of a bottleneck structure composed of multiple convolutional and pooling layers, while the decoder uses a symmetrical deconvolutional structure to reconstruct features. The model has a low overall parameter count and can run directly on ARM architecture edge control boards without requiring external graphics computing units. Single-frame inference latency meets real-time monitoring requirements.

[0035] The three-channel reconstruction error joint verification unit executes a cross-judgment rule: a single signal reconstruction error exceeding the limit is not considered a tool fault; tool damage is only confirmed when at least two of the vibration, acoustic emission, and power signals simultaneously exceed the corresponding level's reconstruction error threshold. The three-channel reconstruction error joint verification unit is configured with graded error thresholds corresponding one-to-one with the five levels of damage; during judgment, matching is performed level by level, and the corresponding damage level result is output only when the reconstruction errors of at least two signals simultaneously exceed the same level threshold.

[0036] The multi-level hierarchical linkage alarm module is connected to the identification module and is used to pre-configure the threshold range corresponding to multi-level tool damage. After confirming the fault through multiple consecutive frames of anti-shake, it performs linkage control of data retention, audible and visual warning and machine tool feed pause / spindle emergency stop according to the damage level. The multi-level hierarchical linkage alarm module is connected to the identification module and is pre-configured with five levels of tool damage corresponding to reconstruction error threshold ranges, which in turn correspond to Level 1 normal cutting, Level 2 slight uniform wear, Level 3 micro-chipping warning, Level 4 severe chipping, and Level 5 complete tool breakage, wherein: The module has built-in continuous multi-frame anti-shake judgment logic. It is necessary to determine that the fault is valid only if multiple consecutive frames are judged to be of the same damage level, and filter out instantaneous random interference. It also performs differentiated linkage according to the damage level: Level 2 slight wear only stores data locally, Level 3 micro-chipping triggers local audible and visual warning and retains the original waveform, and Level 4 and Level 5 damage immediately sends a pause or emergency stop command to the machine tool control system. The multi-level hierarchical linkage alarm module distinguishes the damage level by using lights of different colors and flashing frequencies and buzzers of different rhythms. In the fourth and fifth damage states, in addition to the audible and visual alarms, it simultaneously sends feed pause and spindle braking commands to the machine tool control system through dual links of switch quantity and industrial bus to lock the cutting action and prevent the fault from escalating.

[0037] The machine tool process interaction and synchronization module is connected to the other modules respectively. It is used to store the exclusive baseline templates corresponding to various tools, input tool and cutting process parameters, support manual baseline reset when changing tools and materials, and export fault detection log data, so as to realize the synchronous adaptation between the detector and the machine tool process.

[0038] The machine tool process interaction and synchronization module is connected to the other modules respectively, and has a built-in tool process database, parameter input unit, baseline manual reset unit, and fault log export unit, wherein: The tooling process database stores exclusive baseline templates for tools of different specifications and materials; The parameter input unit is used to input tool specifications, number of teeth, workpiece material, and cutting process parameters; The baseline manual reset unit supports clearing historical baselines and restarting the quick learning process with one click after tool or material change; The fault log export unit supports exporting the original signals, reconstructed error curves, and corresponding process parameters during abnormal periods for process review and optimization.

[0039] like Figure 2 As shown, the present invention also discloses a high-sensitivity broken knife identification method, comprising the following steps: Step S1, Process initialization configuration: Input tool parameters, workpiece material and cutting process parameters through the machine tool process interaction synchronization module, bind the tool and dedicated baseline storage space, and complete sensor zero-point calibration and machine tool communication docking. Step S2, Multi-source high-precision synchronous acquisition: After the machine tool starts cutting, the three-axis vibration, acoustic emission and power torque signals are acquired synchronously, the timing is aligned by a unified timestamp, and the raw sampling data is output. Step S3, Multi-stage Coupling Preprocessing Noise Reduction: Adaptively switch the filter frequency band according to the current speed and tool parameters, dynamically adjust the signal gain, filter out power frequency electromagnetic interference, identify and shield the transient intervals of no-load feed and cut-in / cut-out, and only output the effective signal of steady-state cutting. Step S4: Segmented fast and slow dual-mode baseline self-learning: Identify steady-state cutting ranges by analyzing signal fluctuation characteristics; When starting a new tool or the first piece of production change, a rapid learning mode is activated to generate an initial multi-dimensional benchmark feature vector based on multi-frame steady-state data. The continuous steady-state processing is switched to a slow iteration mode, and the baseline is updated frame by frame with low weight through a fixed-length sliding window. When a significant fluctuation in process parameters is detected, the baseline is frozen, and iteration resumes after stabilization. The reference amplitude is corrected in real time by combining the spindle temperature to compensate for temperature drift. In step S4, if steady-state cutting data is not identified for a preset time, the baseline iteration update is temporarily suspended and the learning process is automatically restarted after stable cutting is restored.

[0040] Step S5: Multi-dimensional time-frequency joint feature extraction: Extract time-domain statistical features, frequency band energy features and time-frequency transient features from the three steady-state effective signals respectively, and splice them to form three independent real-time feature vectors; Step S6, Multi-source Cross-validation Judgment: Input the three real-time feature vectors into the lightweight convolutional autoencoder respectively, calculate the reconstruction error of each relative to the baseline feature vector; execute the cross-validation rule for at least two synchronous over-limits, and match the five-level damage level with the error threshold range; and output the final judgment result after confirmation by multiple consecutive frames of anti-shake. In step S6, the lightweight convolutional autoencoder only uses normal steady-state cutting samples to complete offline pre-training, and does not input any tool breakage or chipping fault samples during the training process; in the on-site machining stage, only inference calculation is performed, and there is no need to perform secondary model fine-tuning for new tools and new materials.

[0041] Step S7, Five-level hierarchical linkage processing: Level 1 Normal Cutting: Continuous monitoring, no additional output; Level 2 Minor Wear: Locally archived data, does not trigger alarms or shutdown; Level 3 micro-chipping: Triggers audible and visual warnings, keeps the machine tool processing, and synchronously stores the original data of the entire channel during the fault period; Level 4 severe chipping, Level 5 complete tool breakage: Immediately issue feed pause and spindle brake commands to the machine tool, continuously sound and light alarms, and lock the machining process; Step S8, Fault Data Archiving: The judgment results of all damage events, original waveforms, reconstructed error curves, and corresponding process parameters are automatically stored in local storage, and can be exported for tool life analysis and cutting process optimization.

[0042] It is worth mentioning that the technical features such as cutting tools involved in this patent application should be regarded as prior art. The specific structure, working principle, and possible control methods and spatial arrangement of these technical features can be adopted using conventional choices in the field, and should not be regarded as the inventive point of this patent. This patent will not be further elaborated in detail.

[0043] For those skilled in the art, modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the protection scope of this invention.

Claims

1. A high-sensitivity broken tool detection system, characterized in that, It includes a multi-source synchronous acquisition module, a multi-stage coupling preprocessing module, a segmented fast and slow dual-mode adaptive baseline self-learning module, an identification module, a multi-level hierarchical linkage alarm module, and a machine tool process interaction synchronization module, among which: The multi-source synchronous acquisition module includes a three-axis MEMS vibration sensing unit, a wideband acoustic emission (AE) sensing unit, a CNC system power and torque acquisition unit, and a high-precision timestamp alignment unit, which are used to synchronously acquire vibration, acoustic emission, and power and torque signals during the cutting process and unify the timing reference of the three signals. The multi-stage coupling preprocessing module is used to perform adaptive bandpass filtering, dynamic gain adjustment, power frequency electromagnetic interference filtering, and cutting interval screening on the acquired multiple raw signals, eliminating invalid noise and transient interference, and outputting standardized effective characteristic signals for the steady-state cutting stage. The segmented fast and slow dual-mode adaptive baseline self-learning module is used to identify the steady-state cutting range. It adopts a dual-mode mechanism of rapid construction of new tools and slow iteration of continuous machining to generate and update multi-dimensional benchmark feature vectors. At the same time, it freezes the baseline when the working conditions change suddenly and combines spindle temperature compensation to eliminate the baseline offset caused by changes in working conditions and environment. The recognition module is used to extract the multi-dimensional time-frequency features of the three signals. After the reconstruction error is calculated by a lightweight convolutional autoencoder trained only with normal cutting samples, the cross-validation rule of at least two signals synchronously exceeding the limit is executed, and the tool damage level judgment result is output. The multi-level hierarchical linkage alarm module is used to pre-configure threshold ranges corresponding to multi-level tool damage. After confirming the fault through multiple consecutive frames of anti-shake, it performs linkage control of data retention, audible and visual warning, and machine tool feed pause / spindle emergency stop according to the damage level. The machine tool process interaction and synchronization module is used to store the exclusive baseline templates corresponding to various tools, input tool and cutting process parameters, support manual baseline reset when changing tools and materials, and export fault detection log data, so as to realize the synchronous adaptation between the testing instrument and the machine tool process.

2. The high-sensitivity broken tool detection system according to claim 1, characterized in that, For the multi-source synchronous acquisition module: The triaxial MEMS vibration sensing unit is fixed to the spindle housing of the CNC machine tool and is used to collect the X / Y / Z three-axis vibration timing signals during the cutting process. The wideband acoustic emission (AE) sensing unit is fitted to the front end face of the spindle and is used to capture high-frequency stress wave signals generated by the propagation of microcracks in the tool, tool tip breakage, and instantaneous fracture. The power and torque acquisition unit of the CNC system is connected to the CNC system of the machine tool via an industrial bus to synchronously read the spindle output power, real-time torque, spindle speed, and feed rate process data. The high-precision timestamp alignment unit adds a unified timing reference to the three heterogeneous signals, eliminating feature misalignment caused by the sampling time difference of multiple channels.

3. The high-sensitivity broken tool detection system according to claim 2, characterized in that, The multi-stage coupling preprocessing module includes an adaptive segmented bandpass filter unit, a dynamically controlled gain unit, an electromagnetic noise notch filter unit, and a cutting interval screening unit, wherein: The adaptive segmented bandpass filter unit dynamically matches the filter passband corresponding to the tooth passing frequency based on the current spindle speed and the number of tool teeth, separating the cutting characteristic frequency band from the machine tool body motion and cutting fluid impact interference frequency band. The dynamic programmable gain unit automatically adjusts the amplification factor according to the real-time signal-to-noise ratio of the signal, increasing the gain for weak breakdown characteristic signals and decreasing the gain for strong cutting steady-state signals. The electromagnetic noise notch filter unit is used to filter out the power frequency electromagnetic interference generated by the machine tool frequency converter. The cutting interval screening unit identifies and marks the transient intervals of empty feed, entry, and exit based on the linkage between signal fluctuation variance and process parameters, and retains only the effective steady-state cutting data, which are then sent to the segmented fast and slow dual-mode adaptive baseline self-learning module and the identification module respectively.

4. The high-sensitivity broken tool detection system according to claim 3, characterized in that, The segmented fast and slow dual-mode adaptive baseline self-learning module is connected to the multi-stage coupled preprocessing module, and includes a cutting interval identification sub-unit, an initial baseline fast construction sub-unit, a sliding window slow iteration sub-unit, a working condition change baseline freezing sub-unit, and a temperature drift compensation sub-unit, wherein: The cutting interval identification subunit divides machining intervals into four categories—no-load, entry, steady-state cutting, and exit—based on signal fluctuation characteristics, selecting only steady-state cutting data for baseline updates. The initial baseline rapid construction subunit, for new tools or production changeover conditions, rapidly generates a multi-dimensional reference feature vector based on the statistical feature distribution of steady-state cutting signals within a preset number of frames. The sliding window slow iteration subunit uses a fixed-length sliding window in continuous steady-state machining, updating the reference feature vector frame by frame with a low weighting to adapt to the gradual normal wear of the tool. The working condition abrupt change baseline freezing subunit pauses baseline updates when it detects that the spindle speed or feed rate fluctuations exceed a preset range, resuming iteration only after the process parameters have stabilized for a preset duration, avoiding abnormal data contaminating the reference template. The temperature drift compensation subunit collects the spindle housing temperature in real time and adjusts the reference feature amplitude through a temperature drift correction model to eliminate the overall baseline offset caused by equipment temperature rise.

5. The high-sensitivity broken tool detection system according to claim 4, characterized in that, The baseline iteration control logic of the segmented fast and slow dual-mode adaptive baseline self-learning module is as follows: The cutting interval identification subunit calculates the signal variance within the sliding window in real time. If the variance is lower than the no-load threshold, it is determined to be an empty stroke. The variance stable interval is marked as steady-state cutting. When a new tool is loaded for the first time or the workpiece material is changed, the fast learning mode is activated to collect the statistical mean and fluctuation range of multiple frames of steady-state cutting data and generate an initial multi-dimensional benchmark feature vector. When the same tool is continuously machining in a steady state, it switches to slow iteration mode. For each new frame of valid signal, the baseline feature vector is updated only with low weight, and the weight of old samples decays synchronously. When the machine tool process interaction synchronization module reads that the spindle speed and feed rate fluctuations exceed the preset range, the working condition sudden change baseline freezing subunit locks the reference template, pauses iteration until the process stabilizes for a preset duration, and then unlocks. The temperature drift compensation subunit reads the spindle temperature sampling value in real time, and corrects the reference characteristic amplitude once for each preset temperature change, eliminating the baseline offset caused by temperature drift.

6. The high-sensitivity broken tool detection system according to claim 5, characterized in that, The recognition module is connected to a multi-stage coupled preprocessing module and a segmented fast-slow dual-mode adaptive baseline self-learning module, and includes a multi-dimensional time-frequency joint feature extraction unit, a lightweight convolutional autoencoder model, and a three-way reconstruction error joint verification unit, wherein: The multi-dimensional time-frequency joint feature extraction unit extracts time-domain statistical features, frequency band energy features and time-frequency transient features from the three preprocessed signals, and splices them together to form three independent multi-dimensional feature vectors. The lightweight convolutional autoencoder model completes unsupervised training using only normal steady-state cutting samples, without the need to label broken or chipped tool fault samples. It adopts a symmetrical lightweight convolutional structure of encoder and decoder, and characterizes the degree to which the tool state deviates from normal cutting by calculating the reconstruction error between the real-time feature vector and the baseline feature vector. The three-way reconstruction error joint verification unit executes a cross-judgment rule: a single signal reconstruction error exceeding the limit is not judged as a tool fault; tool damage is only confirmed when at least two of the three signals of vibration, acoustic emission, and power synchronously exceed the corresponding level of reconstruction error threshold.

7. The high-sensitivity broken tool detection system according to claim 6, characterized in that, The multi-level hierarchical linkage alarm module is connected to the identification module and is pre-configured with five levels of tool damage corresponding to reconstruction error threshold ranges, which in turn correspond to Level 1 normal cutting, Level 2 slight uniform wear, Level 3 micro-chipping warning, Level 4 severe chipping, and Level 5 complete tool breakage, wherein: The module has built-in continuous multi-frame anti-shake judgment logic. It is necessary to determine that the fault is valid only if multiple consecutive frames are judged to be of the same damage level, and to filter out instantaneous random interference. It also performs differentiated linkage according to the damage level: Level 2 slight wear only stores data locally, Level 3 micro-chipping triggers local audible and visual warning and retains the original waveform, and Level 4 and Level 5 damage immediately sends a pause or emergency stop command to the machine tool control system.

8. The high-sensitivity broken tool detection system according to claim 7, characterized in that, The machine tool process interaction and synchronization module is connected to the other modules respectively, and has a built-in tool process database, parameter input unit, baseline manual reset unit, and fault log export unit, wherein: The tooling process database stores exclusive baseline templates for tools of different specifications and materials; The parameter input unit is used to input tool specifications, number of teeth, workpiece material, and cutting process parameters; The baseline manual reset unit supports clearing historical baselines and restarting the quick learning process with one click after tool or material change; The fault log export unit supports exporting the original signals, reconstructed error curves, and corresponding process parameters during abnormal periods for process review and optimization.

9. A high-sensitivity method for identifying broken blades, characterized in that, The high-sensitivity broken blade detection system according to any one of claims 1 to 8 comprises the following steps: Step S1, Process initialization configuration: Input tool parameters, workpiece material and cutting process parameters through the machine tool process interaction synchronization module, bind the tool and dedicated baseline storage space, and complete sensor zero-point calibration and machine tool communication docking. Step S2, Multi-source high-precision synchronous acquisition: After the machine tool starts cutting, the three-axis vibration, acoustic emission and power torque signals are acquired synchronously, the timing is aligned by a unified timestamp, and the raw sampling data is output. Step S3, Multi-stage Coupling Preprocessing Noise Reduction: Adaptively switch the filter frequency band according to the current speed and tool parameters, dynamically adjust the signal gain, filter out power frequency electromagnetic interference, identify and shield the transient intervals of no-load feed and cut-in / cut-out, and only output the effective signal of steady-state cutting. Step S4: Segmented fast and slow dual-mode baseline self-learning: Identify steady-state cutting ranges by analyzing signal fluctuation characteristics; When starting a new tool or the first piece of production change, a rapid learning mode is activated to generate an initial multi-dimensional benchmark feature vector based on multi-frame steady-state data. The continuous steady-state processing is switched to a slow iteration mode, and the baseline is updated frame by frame with low weight through a fixed-length sliding window. When a significant fluctuation in process parameters is detected, the baseline is frozen, and iteration resumes after stabilization. The reference amplitude is corrected in real time based on the spindle temperature to compensate for temperature drift. Step S5: Multi-dimensional time-frequency joint feature extraction: Extract time-domain statistical features, frequency band energy features and time-frequency transient features from the three steady-state effective signals respectively, and splice them to form three independent real-time feature vectors; Step S6, Multi-source cross-validation determination: Input the three real-time feature vectors into the lightweight convolutional autoencoder respectively, calculate the reconstruction error of each relative to the baseline feature vector; execute the cross-validation rule for at least two synchronous over-limits, match the five-level damage level with the error threshold range; and output the final determination result after confirmation by multiple consecutive frames of image stabilization. Step S7, Five-level hierarchical linkage processing: Level 1 Normal Cutting: Continuous monitoring, no additional output; Level 2 Minor Wear: Locally archived data, does not trigger alarms or shutdown; Level 3 micro-chipping: Triggers audible and visual warnings, keeps the machine tool processing, and synchronously stores the original data of the entire channel during the fault period; Level 4 severe chipping, Level 5 complete tool breakage: Immediately issue feed pause and spindle brake commands to the machine tool, continuously sound and light alarms, and lock the machining process; Step S8, Fault Data Archiving: The judgment results of all damage events, original waveforms, reconstructed error curves, and corresponding process parameters are automatically stored in local storage, and can be exported for tool life analysis and cutting process optimization.

Images