Uncertainty reasoning state scoring and early warning system of gear machining equipment

Abstract

The invention discloses an uncertainty reasoning state scoring and early warning system for gear machining equipment, and relates to the technical field of gear machining monitoring. Data are collected in real time through an industrial Ethernet, time synchronization is realized based on an IEEE1588PTP protocol, and a data cleaning sub-module is designed to repair an abnormal value and keep a record; the method comprises the following steps of: extracting 128-dimensional potential features of a vibration signal by adopting a variational auto-encoder, fusing a temperature drift coefficient and a current harmonic component, quantifying feature uncertainty through Monte Carlo dropout, and marking low-reliability features; and constructing a Bayesian network containing 5 state nodes and 12 observation nodes, calculating a posterior probability through a Markov chain Monte Carlo algorithm, dynamically adjusting a conditional probability table, and fusing multi-feature scores. According to the method, the state evaluation accuracy and the early warning timeliness are improved, the resource consumption is optimized, full-link monitoring is realized, the fault handling efficiency is enhanced, and reliable support is provided for precision manufacturing.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of gear machining detection, and particularly relates to an uncertainty reasoning state scoring and early warning system of a gear machining device. BACKGROUND

[0002] As core equipment for precision manufacturing, the running state of a gear machining device directly determines the precision and service life of a gear product. At present, the device state monitoring in the gear machining process mainly relies on single sensor data or manual inspection, which has significant limitations. Single sensor can only capture local features, such as vibration sensor focusing on spindle running fluctuations and temperature sensor only reflecting heat dissipation state, which is difficult to fully depict the overall health condition of the device. Manual inspection is limited by long cycle and strong subjectivity, and often can only intervene after the failure occurs, resulting in production interruption and cost increase.

[0003] Existing state evaluation systems mostly use deterministic models, ignoring uncertain factors such as sensor noise and environmental interference, resulting in a large deviation between the scoring results and the actual state. For example, vibration signals are prone to burrs due to electromagnetic interference during high-speed cutting, which may be misjudged as tool wear if directly used for state judgment; temperature drift will cause the current sensor data to deviate, affecting the accurate evaluation of motor load. In addition, the early warning threshold is mostly a fixed value, which cannot adapt to dynamic scenarios such as device aging and processing material changes, resulting in either frequent false alarms interfering with production or a high threshold causing missed reports and causing serious failures.

[0004] The lack of cross-parameter correlation analysis further restricts the early warning accuracy. Tool wear is not only reflected in the change of cutting force, but also has a strong correlation with spindle vibration frequency and feed speed fluctuation. However, the existing system does not establish a multi-parameter correlation model, making it difficult to identify early failures from a global perspective. At the same time, the mapping relationship between processing precision and device state is fragmented, and it is impossible to predict quality problems such as gear tooth thickness deviation through state scoring, resulting in the frequent phenomenon of "normal device but unqualified product", which seriously affects production efficiency and product qualification rate. SUMMARY

[0005] The uncertainty reasoning state scoring and early warning system of a gear machining device is proposed to solve the problems mentioned in the prior art.

[0006] To achieve the above purpose, the present application adopts the following technical scheme: an uncertainty reasoning state scoring and early warning system of a gear machining device, comprising:

[0007] Information acquisition module: deploy vibration sensor, temperature sensor, spindle current sensor and acoustic sensor to realize real-time acquisition of equipment operation data; adopt timestamp synchronization mechanism, align different sources of data based on IEEE1588 PTP protocol; data cleaning module identifies jump values by 3 times standard deviation rule, repairs the identified abnormal points by linear interpolation of 5 effective values before and after, and retains the original abnormal record;

[0008] Uncertainty feature extraction module: adopt variational autoencoder to construct feature space, including encoder and decoder, extract 128-dimensional features from vibration signal, and fuse temperature drift coefficient and current harmonic component; temperature drift coefficient is the parameter deviation rate corresponding to each °C; adopt Monte Carlo dropout method to quantify feature uncertainty;

[0009] State reasoning and scoring module: construct Bayesian network model containing 5 state nodes, namely clamping, cutting, feeding, detection and unloading, and 12 observation nodes corresponding to each sensor feature; calculate the posterior probability distribution of each state by Markov chain Monte Carlo algorithm, dynamically adjust the conditional probability table between nodes according to the feature reliability, and integrate vibration, temperature and current feature scores to get the comprehensive health score; the weight coefficient is dynamically allocated according to the real-time feature reliability;

[0010] Dynamic early warning threshold module: build threshold learning model based on historical fault data of equipment, which contains 30 fault records; use industry standard threshold when initializing new equipment; when the equipment runs for 30 days and the processing batch N≥50, start adaptive adjustment, update the threshold value by exponential smoothing method, and the balance coefficient α is 0.7;

[0011] Early warning output module: integrate buzzer, red LED warning light, industrial control bus signal and remote terminal APP push four kinds of early warning methods; activate all channels and output forced shutdown signal.

[0012] Further, it further comprises:

[0013] Tool wear reasoning sub-module: collect cutting force signal, perform cross-correlation analysis on cutting force signal and vibration signal, use 512-point sliding window, extract 10 wear features including peak factor and kurtosis; construct LSTM network to predict wear trend in next 5 minutes, network input is 3-minute feature sequence, and output is wear prediction value; use Huber loss function to train the model;

[0014] The processing precision correlation submodule: through the laser diameter measuring instrument to collect gear tooth thickness data, the correlation matrix of tooth thickness deviation and equipment running state is constructed, the tooth thickness deviation is the difference between the measured value and the design value, the Pearson correlation coefficient r of the state characteristics and the precision index is calculated, the features with |r|≥0.7 are marked as strong correlation features, and the correlation between the feed speed fluctuation and the tooth thickness deviation is marked as a strong correlation feature.

[0015] The uncertainty quantification optimization submodule: the D-S evidence theory is introduced to process information conflict, and the basic probability assignment function m(A) is defined, where A is a state proposition. The state judgment results of vibration and temperature sensors are synthesized, and when the sensor data conflict occurs, the probability distribution is corrected through the synthesis rule;

[0016] The adaptive sampling adjustment submodule: the acquisition frequency is dynamically adjusted according to the current state score. For the intermediate state, the sampling frequency f is smoothly transitioned according to the linear formula f=10+90×(1-S / 100), where f is the sampling frequency and S is the current state score.

[0017] Further, the comprehensive health score calculation formula of the state reasoning and scoring module is: S=ω1·S+ω2·S t +ω3·S, wherein S is the comprehensive health score, S is the vibration feature score, S t is the temperature feature score, S is the current feature score, ω1, ω2, ω3 are weight coefficients and satisfy ω1+ω2+ω3=1, and the weight value is dynamically allocated according to the reliability score of each feature;

[0018] The threshold updating formula of the dynamic early warning threshold module is: T k+1 =α·T k +(1-α)·μ k , wherein T k+1 is the updated threshold of the k+1 batch, T k is the current threshold of the k batch, μ k is the state score mean of the k batch processing, and α is a smoothing coefficient, which realizes the slow adaptive adjustment of the threshold with the long-term running state of the equipment.

[0019] Further, the wear amount prediction model of the tool wear reasoning submodule adopts Huber loss function: when , when , wherein L is the loss value, is the predicted wear amount, y is the actual wear amount, and δ is the threshold, which realizes the reduction of the interference of abnormal values on model training.

[0020] Further, the evidence synthesis rule of the uncertainty quantification optimization submodule is: Wherein m1, m2 are the basic probability assignment functions of two sensors respectively, is the evidence combination operator, A, B, C are propositions in the recognition framework.

[0021] Further, the time synchronization mechanism of the information acquisition module adopts a master-slave clock architecture, the master clock is a device controller, and the slave clock is each sensor node.

[0022] Further, the VAE model reconstruction loss of the uncertainty feature extraction module is: Loss=β·MSE(x,)+(1-β)·KL(q(z|x)||p(z)), wherein Loss is the reconstruction loss value, β is a balance coefficient, and the value is 0.7, MSE is a mean square error function, x is an input original feature vector, is a reconstructed feature vector output by the VAE decoder, KL is a KL divergence function, q(z|x) is an approximate posterior probability distribution output by the encoder, and p(z) is a preset prior probability distribution.

[0023] The early warning level determination rule of the early warning output module is: when the comprehensive health score S is greater than or equal to 80, it is a normal state without early warning; when 60 is less than or equal to S and less than 80, it is a first-level early warning corresponding to a slight abnormality; when 40 is less than or equal to S and less than 60, it is a second-level early warning corresponding to a moderate abnormality; when 20 is less than or equal to S and less than 40, it is a third-level early warning corresponding to a serious abnormality; and when S is less than 20, it is a fourth-level early warning corresponding to a critical state, and different response mechanisms correspond to different levels of early warning.

[0024] Compared with the prior art, the present application has the following advantages:

[0025] In terms of state evaluation accuracy, the system innovatively introduces an uncertainty reasoning mechanism, quantifies the noise interference of sensor data through a variational autoencoder, and combines D-S evidence theory to fuse multi-source information conflicts, effectively reducing the risk of single sensor misjudgment. The Bayesian network model dynamically adjusts the feature weight, so that the high reliability parameter occupies a dominant position in the score, solving the problem of ignoring interference factors in traditional deterministic models, and making the state score more consistent with the actual operation of the equipment.

[0026] In terms of early warning timeliness and adaptability, the dynamic threshold module is based on historical data and real-time state to adaptively update, avoiding the problem of fixed threshold inapplicability when the equipment ages or the working condition changes. The multi-level early warning mechanism responds according to the degree of abnormality, and only triggers local reminders for slight abnormalities, and forces shutdown for critical states, which not only reduces unnecessary production interruptions, but also intervenes in time before the failure expands, balancing production continuity and equipment safety.

[0027] In terms of functional integrity, the system integrates tool wear prediction and machining precision correlation analysis, realizing full-link monitoring from equipment state to product quality. The tool wear reasoning sub-module predicts the replacement time in advance, reducing the risk of sudden tool breakage; the machining precision correlation sub-module maps the features to the tooth thickness deviation, so that the state evaluation directly serves the quality control, avoiding the problem of "unqualified products in qualified state".

[0028] In terms of system practicability, the adaptive sampling adjustment mechanism dynamically adjusts the data acquisition frequency according to the state score, reducing the system overhead while ensuring monitoring accuracy. The multi-modal early warning output takes into account both on-site operation and remote management, ensuring that different roles can timely obtain abnormal information and improving fault handling efficiency. Overall, the system greatly improves the intelligent monitoring level of gear machining equipment, providing reliable technical support for precision manufacturing. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 A schematic block diagram of the uncertainty reasoning state scoring and early warning system for the gear machining equipment according to the present application;

[0030] Figure 2 A state evaluation accuracy comparison diagram of the uncertainty reasoning state scoring and early warning system for the gear machining equipment according to the present application under different monitoring modes;

[0031] Figure 3 A schematic diagram of the adaptive adjustment trend of the early warning threshold of the uncertainty reasoning state scoring and early warning system for the gear machining equipment according to the present application. DETAILED DESCRIPTION

[0032] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0033] In the description of the present application, it should be understood that the terms "center", "longitudinal", "transverse", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation of the present application.

[0034] Furthermore, the terms "first", "second", etc. are used only for descriptive purposes and do not connote or imply relative importance or a quantity of the indicated technical features. Thus, features with "first", "second" designations can include one or more of the features implicitly or explicitly. In the description of the present application, the meaning of "a plurality" is two or more, unless otherwise expressly specified. In addition, the terms "mounting", "connected", "connecting" should be interpreted broadly, for example, can be fixedly connected, can be detachably connected, or integrally connected; can be mechanically connected, can be electrically connected; can be directly connected, can be indirectly connected through an intermediate medium, or can be internal communication of two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances, and the present application will be further described in detail below with reference to the drawings.

[0035] Referring to Figures 1 to 3 : An uncertainty reasoning state scoring and early warning system of gear machining equipment, comprising:

[0036] A multi-source heterogeneous information acquisition module: vibration sensors, temperature sensors, main shaft current sensors and acoustic sensors are deployed. The sampling rate of the vibration sensor is set to 2 kHz, the range covers 0 to 50 g, and the resolution reaches 0.001 g; the measurement range of the temperature sensor extends from -20°C to 150°C, and the accuracy is controlled to ±0.5°C; the range of the main shaft current sensor is 0 to 50A, and the accuracy is 0.5%FS; the frequency response range of the acoustic sensor is 20Hz to 20kHz, and the sensitivity is -40dB±3dB. Data transmission is carried out through industrial Ethernet, and the transmission rate is 100Mbps, so as to realize real-time acquisition of equipment operation data. A timestamp synchronization mechanism is adopted, and different sources of data are aligned to 1ms level based on IEEE1588PTP protocol. When the synchronization error exceeds 1ms, the hardware clock calibration is triggered. A data cleaning submodule is designed, and the 3 times standard deviation rule is used to identify jump values, such as vibration signal instantaneous exceeding 10g, current sudden increase of more than 20%, etc. The identified abnormal points are repaired by linear interpolation of 5 effective values before and after, and the original abnormal record is retained at the same time, which is used for subsequent sensor fault diagnosis.

[0037] The uncertainty feature extraction module: a variational autoencoder is used to construct a feature space, which includes an encoder and a decoder. The encoder is a 3-layer fully connected structure, with node numbers of 512, 256, and 128 in sequence. The decoder is also a 3-layer fully connected structure, with node numbers of 128, 256, and 512 in sequence. Through this network, 128-dimensional latent features are extracted from the vibration signal, while the temperature drift coefficient and the current harmonic component are fused. The temperature drift coefficient is the parameter deviation rate corresponding to each °C, and the current harmonic component is the 2-5th harmonic proportion. The Monte Carlo dropout method is used to quantify the feature uncertainty, with a dropout rate of 0.15 and 50 iterations to generate a 95% confidence interval. The vibration feature confidence threshold is set to 0.85. Features below this value, such as vibration components with a signal-to-noise ratio <20 dB, are marked as low reliability features, and their weights are reduced in the subsequent inference process.

[0038] The state inference and scoring module: a Bayesian network model is constructed, which includes 5 key state nodes, namely clamping, cutting, feeding, detection, and unloading, and 12 observation nodes corresponding to each sensor feature. The Markov chain Monte Carlo algorithm is used to calculate the posterior probability distribution of each state, with 1000 iterations and a burn-in period of 500. The conditional probability table between nodes is dynamically adjusted according to the feature reliability. For every 10% decrease in feature reliability, the conditional probability weight is reduced by 15%. The comprehensive health score is fused with the vibration, temperature, and current feature scores, with the weight coefficient dynamically allocated according to the real-time feature reliability to ensure that the weight of high reliability features is increased by 10%-30%.

[0039] The dynamic early warning threshold module: a threshold learning model is constructed based on historical failure data of the equipment, which contains at least 30 typical failure records. When a new device is initialized, industry standard thresholds are used, such as a vibration warning benchmark value of 2.5g for gear machining equipment. After the device has been running for 30 days and the processing batch N≥50, adaptive adjustment is started, and the threshold is updated by the exponential smoothing method, with a balance coefficient α of 0.7 to preserve historical trends and respond to recent changes, avoiding threshold failure due to equipment aging.

[0040] Multi-modal early warning output module: integrate 85dB buzzer, red LED warning light, industrial control bus signal and remote terminal APP push four kinds of early warning mode. Among them, the continuous ringing period of 85dB buzzer is 2s, the red LED warning light flashes at a frequency of 10Hz, the industrial control bus signal adopts Profinet protocol, outputs switching value, and the remote terminal APP push delay is ≤3s. According to the comprehensive score, four levels of early warning are divided: the first level of early warning corresponds to 60≤S<80, only triggering local sound and light alarm; the second level of early warning corresponds to 40≤S<60, increasing the bus signal output; the third level of early warning corresponds to 20≤S<40, synchronously pushing the APP early warning; the fourth level of early warning corresponds to S<20, activating all channels and outputting the forced shutdown signal, to ensure that the risk is disposed in time.

[0041] In the present application, further comprising:

[0042] Tool wear inference sub-module and machining precision correlation sub-module: the tool wear inference sub-module is used for real-time monitoring of tool wear state and prediction of wear trend, the sub-module collects cutting force signals, the sampling frequency is set to 1kHz, the signal range covers 0 to 10kN, and the measurement accuracy is controlled within 1%FS. The collected cutting force signals and the vibration signals obtained by the vibration sensor are subjected to cross-correlation analysis, the analysis process adopts a 512-point sliding window, and 10 characteristic parameters related to tool wear are extracted through the analysis, including peak factor, kurtosis, pulse factor, margin factor, waveform factor, skewness, variance, root mean square, peak value and mean value. An LSTM network is constructed for predicting tool wear trend in the next 5 minutes, the network includes 2 layers of hidden layers, each layer of hidden layer is provided with 64 nodes, and a dropout rate of 0.2 is set to prevent overfitting. The input of the network is a feature sequence of the last 3 minutes, and the output is a predicted value of tool wear amount. The prediction model is trained by using a Huber loss function, and the expression of the function is: when , when , wherein L is a loss value, is a predicted tool wear amount, in mm, y is an actually measured tool wear amount, in mm, and δ is a threshold value, which is 0.01mm. When the predicted tool wear amount exceeds 70% of the total life of the tool, for example, for a carbide tool, when the predicted wear amount is greater than or equal to 0.3mm, the sub-module triggers a preventive maintenance prompt, and displays the estimated number of remaining workpieces on the operation interface of the device.

[0043] The processing precision correlation submodule is used to establish the correlation between the device running state and the gear processing precision, the tooth thickness data of the gear is collected by a laser diameter measuring instrument, the measurement precision of the laser diameter measuring instrument is ±0.001 mm, and the sampling interval is set to 1 s. The tooth thickness deviation, that is, the difference between the measured tooth thickness value and the designed tooth thickness value, is calculated, and the correlation matrix between the tooth thickness deviation and the device running state features is constructed, the dimension of the matrix is 128*5. The Pearson correlation coefficient r of each state feature and the precision index of the tooth thickness deviation is calculated, and the calculation formula is: r = Σ [(xi-μx)(yi-μy)] / √[Σ(xi-μx) 2 Σ(yi-μy) 2 ], wherein xi is the i th state feature value, μx is the mean value of the state feature value, yi is the i th tooth thickness deviation value, and μy is the mean value of the tooth thickness deviation value. The state features with |r|≥0.7 are marked as the features strongly correlated with the processing precision, for example, the correlation feature of the feed speed fluctuation and the tooth thickness deviation, in the comprehensive health score calculation, the weight of these strongly correlated features is increased by 20%, so as to enhance the correlation between the device state evaluation and the actual processing quality.

[0044] In the present application, the following are also included:

[0045] The uncertainty quantification optimization submodule and the adaptive sampling adjustment submodule: the uncertainty quantification optimization submodule is used to optimize the uncertainty quantification results of the multi-source information, and improve the accuracy of state judgment, the submodule introduces D-S evidence theory to process the conflict between multi-source sensor information. Define the basic probability assignment function m(A), wherein A is a state proposition in the recognition framework, for example, "device normal", "clamping abnormal", "cutting abnormal" and the like. The state judgment results of the vibration sensor and the temperature sensor are synthesized, when the sensor data conflicts, for example, the vibration sensor prompts the device to be abnormal, but the temperature sensor shows that the device is normal, the probability distribution is corrected through the evidence synthesis rule. The expression of the evidence synthesis rule is:

[0046] Wherein m1 and m2 are the basic probability assignment functions of the vibration sensor and the temperature sensor respectively, ⊕ represents the evidence synthesis operator, A, B and C are propositions in the recognition framework, through the rule, the conflicting evidence can be effectively fused, and the state misjudgment rate caused by single sensor failure is controlled to be less than 5%.

[0047] The adaptive sampling adjustment submodule is used for dynamically adjusting the data acquisition frequency according to the current state of the equipment, so as to balance the integrity of data acquisition and system resource consumption. The submodule dynamically adjusts the acquisition frequency f of each sensor according to the current state score S output by the state reasoning and scoring submodule. When the state score S is greater than or equal to 80 points, a basic sampling rate of 10 Hz is adopted; when the state score S is less than 60 points, a high-frequency sampling rate of 100 Hz is adopted; for the intermediate state of 60≤S<80, the sampling frequency is smoothly transitioned according to the linear formula f=10+90×(1-S / 100), wherein f is the sampling frequency, the unit is Hz, S is the current state score, and the value range is 0-100. At the same time, in order to avoid the sampling frequency exceeding the capability range of the hardware equipment due to the extreme score, the upper limit of the sampling frequency is set to 100 Hz, and the lower limit is set to 10 Hz. Through this dynamic adjustment mechanism, the data acquisition amount is reduced when the equipment state is good to reduce the system overhead, and the data acquisition amount is increased when the equipment state is poor to ensure the accuracy of the equipment state monitoring.

[0048] In the present application, the comprehensive health score of the state reasoning and scoring module is calculated by weighted summation, and the calculation formula is: S=ω1·S+ω2·S t +ω3·S, wherein S is the comprehensive health score, the value range is 0-100; S is the vibration feature score, which is obtained after extracting features from the data collected by the vibration sensor, the value range is 0-100; S t is the temperature feature score, which is obtained after extracting features from the data collected by the temperature sensor, the value range is 0-100; S is the current feature score, which is obtained after extracting features from the data collected by the main shaft current sensor, the value range is 0-100; ω1, ω2, ω3 are weight coefficients of the vibration feature score, the temperature feature score and the current feature score respectively, and satisfy ω1+ω2+ω3=1. The values of the weight coefficients are dynamically allocated according to the reliability scores of the features. The higher the reliability score of the feature, the larger the corresponding weight coefficient. Specifically, the corresponding weight increases by 5%-10% for every 10% increase in feature reliability. For example, when the reliability score of the vibration feature increases by 20% from the initial value, ω1 increases by 10%-20% based on the original value, so as to ensure that the high-reliability feature occupies a more important position in the comprehensive health score and improves the accuracy of the score.

[0049] In the present application, the threshold updating of the dynamic early warning threshold module adopts the exponential smoothing method, and the updating formula is: T k+1 =α·T k +(1-α)·μ k , wherein T k+1 is the updated early warning threshold in the k+1 batch processing; T k is the current early warning threshold in the k batch processing; μk The average state score output by the state reasoning and scoring module during the k-th batch of processing is denoted as α; α is a smoothing coefficient, ranging from 0.6 to 0.8, with a default value of 0.7. This formula ensures that the updated warning threshold retains the trend of historical thresholds while responding to recent state changes during processing. For example, when equipment ages due to long-term operation, causing the average state score to gradually decrease, the warning threshold can be lowered slowly. This avoids delayed or false warnings due to a fixed threshold, ensuring that the threshold can adaptively adjust with the long-term operating status of the equipment, improving the accuracy and timeliness of warnings.

[0050] In this invention, the wear prediction model of the tool wear inference submodule employs the Huber loss function during training to reduce the interference of outliers on model training. The expression of this function is: when hour, when hour, Where L is the loss value, used to measure the deviation between the predicted value and the actual value; y represents the tool wear predicted by the model, in mm; y represents the tool wear obtained through actual measurement, in mm; δ is the threshold value, set to 0.01 mm. During model training, when the absolute value of the difference between the predicted and actual wear is less than or equal to 0.01 mm, a squared loss function is used to more accurately penalize smaller errors; when the absolute value of the difference is greater than 0.01 mm, a linear loss function is used to avoid large outliers (such as wear jumps caused by measurement errors) from having an excessive impact on model parameter updates, enabling the model to learn the tool wear pattern more robustly and improving the accuracy of wear prediction.

[0051] In this invention, the uncertainty quantification optimization submodule uses the evidence synthesis rule in DS evidence theory to fuse the state judgment results of multi-source sensors. The expression of this rule is: Wherein, m1 is the basic probability assignment function of the vibration sensor to each state proposition, m2 is the basic probability assignment function of the temperature sensor to each state proposition, is the evidence combination operator, A, B, C are all state propositions in the recognition framework, such as "normal equipment", "abnormal cutting", "abnormal feeding", etc.; the numerator part represents the sum of the basic probability assignment products of all intersections of B and C as A; the denominator part is the normalization factor, which is 1 minus the sum of the basic probability assignment products of all intersections of B and C as the empty set. Through the evidence combination rule, the state judgment information from different sensors can be effectively fused, and when the judgment results of different sensors conflict, such as the vibration sensor considering that the probability of "abnormal cutting" is 0.8, and the temperature sensor considering that the probability of "normal cutting" is 0.7, a comprehensive probability distribution result can be obtained through the rule, the influence of single sensor failure or misjudgment on the overall state judgment is reduced, and the robustness of equipment state judgment is improved.

[0052] In the application, the time synchronization mechanism of the multi-source heterogeneous information collection module adopts a master-slave clock architecture to ensure that the data collected by different sensors is consistent in time. The master clock is set as the main controller of the equipment, and its clock accuracy reaches ±1ppm, which can provide a high-precision time reference; each sensor node acts as a slave clock and synchronizes with the master clock through the IEEE1588 PTP protocol. The time calibration period is set to be performed once every 100ms, that is, the main controller sends a time synchronization signal to each sensor node once every 100ms, and each sensor node corrects its clock deviation according to the signal to ensure that the time synchronization error e between all sensor nodes and the main controller is ≤1ms. When the time synchronization error e is detected for three times in succession, the system automatically triggers the restart calibration process of the sensor node, the sensor node restarts and synchronizes with the main controller again, and records the clock deviation value each time, which will be one of the bases for sensor fault diagnosis. If the clock deviation value of a certain sensor is continuously large and cannot be corrected by calibration, it is judged that the sensor may have a hardware fault and needs to be repaired or replaced.

[0053] In the application, the VAE model reconstruction loss of the uncertainty feature extraction module is calculated by the weighted sum of mean square error and KL divergence, and its expression is: Loss=β·MSE(x,

[0054] ) + (1 - β) · KL(q(z|x) || p(z)), wherein, Loss is a reconstruction loss value of the VAE model; β is a balance coefficient, and a value of the balance coefficient is 0.7, and the balance coefficient is used for balancing weights of feature reconstruction accuracy and latent space regularity; MSE is a mean square error function, and is used for calculating a difference between an original feature vector input and a reconstructed feature vector output by the decoder, x is the original feature vector input, and is the reconstructed feature vector output by the VAE decoder; KL is a KL divergence function, and is used for measuring a difference between an approximate posterior probability distribution output by the encoder and a preset prior probability distribution, q(z|x) is the approximate posterior probability distribution output by the encoder, and p(z) is the preset prior probability distribution, and the prior probability distribution adopts a standard normal distribution. The reconstruction loss is calculated through the formula, and in the model training process, the similarity between the reconstructed feature and the original feature can be ensured, and the latent feature distribution output by the encoder is close to the standard normal distribution, the regularity of the feature space is improved, and therefore the uncertainty of the feature is quantified more effectively.

[0055] In the present application, the multi-modal early warning output module determines the early warning level according to the comprehensive health score S output by the state reasoning and scoring module, and triggers the corresponding early warning response mechanism. The specific early warning level determination rule is: when the comprehensive health score S is greater than or equal to 80 points, it is determined that the equipment is in a normal state, at this time the system has no any early warning output; when 60≤S<80 points, it is determined as a first level early warning, corresponding to the equipment appearing slight abnormality, at this time only the local sound and light alarm is triggered, that is, the 85dB buzzer continuously buzzes with a period of 2s, and the red LED warning light flashes with a frequency of 10Hz; when 40≤S<60 points, it is determined as a second level early warning, corresponding to the equipment appearing moderate abnormality, on the basis of the first level early warning, the industrial control bus signal output is increased, the on-off signal is output to the equipment control system through the Profinet protocol, prompting the control system to make corresponding adjustment; when 20≤S<40 points, it is determined as a third level early warning, corresponding to the equipment appearing serious abnormality, on the basis of the second level early warning, the early warning information is pushed to the remote terminal APP synchronously, ensuring that the equipment manager can know the abnormal situation in time, and the push delay control is within 3s; when S<20 points, it is determined as a fourth level early warning, corresponding to the equipment being in a critical state, at this time all early warning channels are activated, and a forced shutdown signal is output, so that the equipment stops running immediately, preventing the fault from further expanding. Different response mechanisms correspond to different levels of early warning, targeted disposal measures can be taken according to the severity of the equipment abnormality, and the safe and stable operation of the equipment is ensured.

[0056] Specific implementation of the uncertainty reasoning state scoring and early warning system of the gear machining equipment.

[0057] Embodiment 1: high-speed gear hobbing machine state monitoring and early warning system

[0058] This embodiment is applied to a certain type of high-speed hobbing machine (maximum machining diameter 500 mm), aiming at typical faults such as clamping deflection and tool wear in gear machining, a multi-source heterogeneous information acquisition module is deployed: a vibration sensor (sampling rate 2 kHz, range 0-50 g) is installed on the spindle box to monitor X / Y / Z three-axis vibration; a temperature sensor (-20℃-150℃) is embedded in the motor shell to collect winding temperature; a current sensor (0-50 A) is connected in series in the main shaft power supply circuit to record current fluctuations; an acoustic sensor (20 Hz-20 kHz) is arranged in the cutting area to capture cutting noise. All sensors are connected to the edge gateway through industrial Ethernet, and 1 ms level time synchronization is realized based on IEEE1588 PTP protocol, and clock calibration is automatically triggered when the synchronization error is more than 1 ms.

[0059] The uncertainty feature extraction module pre-processes the original data: the vibration signal is extracted by a variational autoencoder (encoder 3 layers of full connection, nodes 512-256-128) to extract 128-dimensional features, and the temperature drift coefficient (0.02% parameter deviation per ℃) and 3rd harmonic proportion (current harmonic component) are fused. The feature confidence interval is generated by Monte Carlo dropout (dropout rate 0.15, iteration 50 times), and the vibration feature confidence below 0.85 is marked as low reliability.

[0060] The state reasoning and scoring module constructs a Bayesian network containing 5 state nodes such as clamping and cutting, and calculates the posterior probability by MCMC algorithm (iteration 1000 times). The comprehensive health score is calculated as S=ω1·S+ω2·S t +ω3·S, wherein the vibration score S weight ω1 is dynamically adjusted according to the reliability (the weight increases by 15% when the reliability improves by 20%). The dynamic warning threshold module initializes the industry standard (vibration warning value 2.5g), and updates the threshold after running for 30 days as T k+1 =0.7·T k +0.3·μ k .

[0061] The tool wear reasoning sub-module collects the cutting force signal (1 kHz sampling), and correlates with the vibration signal (512-point window), predicts the wear trend through an LSTM network (2 layers of hidden layers, 64 nodes), and trains it using a Huber loss function (δ=0.01 mm). The machining precision correlation sub-module collects the tooth thickness through a laser diameter gauge (±0.001 mm), calculates the Pearson correlation coefficient r with the state features, and the feature weight is increased by 20% when |r|≥0.7.

[0062] Monitoring mode State evaluation accuracy Early warning timeliness Tool life utilization rate Traditional single sensor Low Lag Low The system of the present application (Embodiment 1) High Timely High

[0063] Traditional single sensor is easily interfered by one-sided information, and the state evaluation is prone to lag behind the actual failure. The system of the application comprehensively captures the device state through multi-source fusion and uncertainty reasoning, and the evaluation accuracy is significantly improved. Dynamic threshold and multi-level warning ensure timely response to abnormalities, and tool wear prediction and precision correlation analysis improve tool utilization and reduce production interruptions caused by sudden failures, balancing processing quality and production efficiency.

[0064] Embodiment 2: Precision gear shaping machine full-process monitoring system

[0065] This embodiment focuses on the precision gear shaping machine (machining precision IT6 level), optimizes the module configuration: adds a grating ruler (resolution 0.1 μm) to collect the displacement of the workbench in the multi-source heterogeneous information collection module, and forms a monitoring network with the original sensor. When cleaning data, the linear interpolation repair is used for the grating ruler jump value (more than 3 times the standard deviation).

[0066] The uncertainty quantification optimization submodule introduces the D-S evidence theory to synthesize the evidence of the state judgment of the vibration and temperature sensors, and the state evaluation is performed according to the The fusion of conflicting information reduces the impact of single sensor failure. The adaptive sampling adjustment submodule dynamically adjusts the frequency according to the score S, 10 Hz sampling when S≥80, 100 Hz sampling when S<60, and the intermediate state transitions according to f=10+90×(1-S / 100).

[0067] The state reasoning and scoring module strengthens the feed state evaluation, the Bayesian network adds a "feed precision" node, and the conditional probability table is dynamically updated with the grating ruler data. The multi-modal warning output module locks the workbench with the grating ruler when the four-level warning is triggered to prevent over-difference machining. The VAE model reconstruction loss is calculated according to Loss=0.7·MSE(x,)+0.3·KL(q(z|x)||p(z)), balancing feature reconstruction and uncertainty quantification.

[0068] Functional index Traditional manual monitoring The system of the present application (Embodiment 2) Comprehensiveness of fault identification One-sided Comprehensive Sampling resource consumption Fixed high consumption Dynamic optimization Quality correlation degree Low High

[0069] Traditional manual monitoring relies on experience and is difficult to cover multi-link failures, and the fixed sampling frequency leads to resource waste; the system of the application realizes full-process state perception through the addition of a grating ruler and evidence synthesis, and the fault identification is more comprehensive. Adaptive sampling reduces invalid data collection and saves system resources. The strong correlation between feed state and precision design makes the state evaluation directly serve the processing quality control, reduces the tooth profile over-difference caused by abnormal device state, and is especially suitable for precision gear production scenes.

[0070] The above merely describes preferred embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can make equivalent replacements or changes within the technical scope disclosed by the present application and according to the technical solutions and inventive concept of the present application, which should be covered within the protection scope of the present application.

Claims

1. An uncertainty reasoning state scoring and early warning system for gear machining equipment, characterized in that, Comprise: Information acquisition module: deploy vibration sensor, temperature sensor, spindle current sensor and sound sensor to realize real-time acquisition of equipment operation data; adopt timestamp synchronization mechanism, align different sources of data based on IEEE1588 PTP protocol; Data cleaning module identifies jump value by 3 times standard deviation rule, repairs the identified abnormal points by linear interpolation of the previous and next 5 effective values, while retaining the original abnormal record; Uncertainty feature extraction module: use variational autoencoder to construct feature space, including encoder and decoder, extract 128-dimensional features from vibration signal, and fuse temperature drift coefficient and current harmonic component; temperature drift coefficient is the parameter deviation rate corresponding to each ℃; use Monte Carlo dropout method to quantify feature uncertainty; State reasoning and scoring module: build Bayesian network model containing 5 state nodes, namely clamping, cutting, feeding, detection and unloading, and 12 observation nodes corresponding to each sensor feature; calculate the posterior probability distribution of each state by Markov chain Monte Carlo algorithm, dynamically adjust the conditional probability table between nodes according to the feature reliability, and integrate vibration, temperature and current feature scores to get the comprehensive health score, with the weight coefficient dynamically allocated according to the real-time feature reliability; Dynamic early warning threshold module: build threshold learning model based on historical fault data of equipment, which contains 30 fault records; use industry standard threshold when initializing new equipment; when the equipment runs for 30 days and the processing batch N is greater than or equal to 50, start adaptive adjustment, update the threshold value by exponential smoothing method, and the balance coefficient α is 0.7; Early warning output module: integrate four early warning methods of buzzer, red LED warning light, industrial control bus signal and remote terminal APP push; activate all channels and output forced shutdown signal.

2. The uncertainty reasoning state scoring and early warning system for gear machining equipment of claim 1, wherein, Also include: Tool wear reasoning sub-module: collect cutting force signal, perform cross-correlation analysis on cutting force signal and vibration signal, use 512-point sliding window to extract 10 wear features, including peak factor and kurtosis; build LSTM network to predict wear trend in next 5 minutes, network input is 3-minute feature sequence, output is wear prediction value; use Huber loss function to train the model; Processing precision correlation sub-module: collect gear tooth thickness data by laser diameter measuring instrument, build the correlation matrix between tooth thickness deviation and equipment running state, tooth thickness deviation is the difference between measured value and design value, calculate the Pearson correlation coefficient r between state features and precision indicators, mark the features with |r|≥0.7 as strong correlation features, including feed speed fluctuation and tooth thickness deviation.

3. The uncertainty reasoning state scoring and early warning system for gear machining equipment of claim 1, wherein, Also include: Uncertainty quantification optimization sub-module: introduce D-S evidence theory to handle information conflict, define basic probability assignment function m(A), where A is a state proposition, synthesize the state judgment results of vibration and temperature sensors when there is conflict between sensor data, and correct the probability assignment through synthesis rules; Adaptive sampling adjustment submodule: dynamically adjust the acquisition frequency according to the current state score, for the intermediate state, according to the linear formula f = 10 + 90 × (1-S / 100) smooth transition, wherein f is the sampling frequency, S is the current state score.

4. The uncertainty reasoning state scoring and early warning system for gear machining equipment of claim 1, wherein, The comprehensive health score calculation formula of the state inference and scoring module is: S=ω1·S+ω2·S+ω3·S, wherein S is the comprehensive health score, S is the vibration feature score, S is the temperature feature score, S is the current feature score, ω1, ω2, and ω3 are weight coefficients and satisfy ω1+ω2+ω3=1, and the weight values are dynamically distributed according to the reliability scores of the features. t t The comprehensive health score calculation formula of the state inference and scoring module is: S=ω1·S+ω2·S+ω3·S, wherein S is the comprehensive health score, S is the vibration feature score, S is the temperature feature score, S is the current feature score, ω1, ω2, and ω3 are weight coefficients and satisfy ω1+ω2+ω3=1, and the weight values are dynamically distributed according to the reliability scores of the features.​ 5. The uncertainty reasoning state scoring and early warning system for gear machining equipment of claim 1, wherein, The threshold updating formula of the dynamic early warning threshold module is: T k+1 = α · T k + (1-α) · μ k , wherein T k+1 is the updated threshold of the k+1 batch, T k is the current threshold of the k batch, μ k is the state score mean of the k batch, and α is a smoothing coefficient, so that the threshold is slowly self-adapted to the long-term running state of the equipment.

6. The uncertainty reasoning state scoring and early warning system for gear machining equipment of claim 2, wherein, The wear amount prediction model of the tool wear reasoning sub-module adopts Huber loss function: when , when δ, wherein L is a loss value, is a predicted wear amount, y is an actual wear amount, and δ is a threshold value, which reduces the interference of outliers on model training.

7. The uncertainty reasoning state scoring and early warning system for gear machining equipment of claim 3, wherein, The evidence combination rule of the uncertainty quantification optimization sub-module is: Wherein m1 and m2 are basic probability assignment functions of two sensors respectively, Denotes an evidence combination operator, and A, B and C are propositions in a recognition framework.

8. The uncertainty reasoning state scoring and early warning system for gear machining equipment of claim 1, wherein, The time synchronization mechanism of the information acquisition module adopts a master-slave clock architecture, the master clock is the device controller, and the slave clock is each sensor node. When e>1ms is detected for three times in succession, the sensor node is triggered to restart the calibration process, and the clock deviation value is recorded for fault diagnosis.

9. The uncertainty reasoning state scoring and early warning system for gear machining equipment of claim 1, wherein, The VAE model reconstruction loss of the uncertainty feature extraction module is: Loss = β·MSE(x,)+(1-β)·KL(q(z|x)||p(z)), wherein Loss is the reconstruction loss value, β is a balance coefficient, the value is 0.7, MSE is a mean square error function, x is an input original feature vector, is a reconstructed feature vector output by a VAE decoder, KL is a KL divergence function, q(z|x) is an encoder output approximate posterior probability distribution, and p(z) is a preset prior probability distribution.

10. The uncertainty reasoning state scoring and early warning system for gear machining equipment of claim 1, wherein, The early warning output module has an early warning level determination rule: when the comprehensive health score S is greater than or equal to 80, it is a normal state, and there is no early warning; when 60 is less than S and S is less than 80, it is a first-level early warning, corresponding to a slight abnormality; When 40 is less than S and S is less than 60, it is a second-level early warning, corresponding to a moderate abnormality; when 20 is less than S and S is less than 40, it is a third-level early warning, corresponding to a serious abnormality; and when S is less than 20, it is a fourth-level early warning, corresponding to a critical state, and each level of early warning corresponds to a different response mechanism.

Images