Fusion model-based shear point prediction method and device

Abstract

The invention relates to the technical field of ultrasonic equipment, and provides a shear point prediction method and device based on a fusion model, and the method comprises the steps: building a first model with better prediction advancement and a second model with better reliability through historical electric signal parameters, and achieving the prediction of a shear point under the condition that a cut object is sheared by an ultrasonic knife. Variable parameters of the real-time electric signals of the ultrasonic knife generator are obtained; obtaining a first output of the first model and a second output of the second model; and obtaining a prediction result of the shear point moment based on a prediction demand of an operator according to the prediction demand, the first output and the second output, and giving a shear point prediction result meeting the demand, so that the situation that the operator observes that the cut object is sheared, resulting in that a cutter head of the ultrasonic knife generates a relatively high temperature, and the shear point prediction result meets the demand is avoided. And the service life of the ultrasonic knife is shortened.

Description

Technical Field

[0001] The present application relates to the technical field of ultrasonic equipment, and in particular to a shear point prediction method and device based on a fusion model. Background Art

[0002] The ultrasonic scalpel transfers electrical energy to the piezoelectric converter through the energy generator, and the piezoelectric converter converts the electrical energy into ultrasonic mechanical energy, thereby obtaining ultrasonic vibrations. The ultrasonic scalpel's blade head amplifies the ultrasonic vibrations to shear the object being cut.

[0003] When cutting the object to be cut, the operator uses observation to determine whether the ultrasonic knife has finished cutting the object to be cut, and cuts off the object to be cut. However, in actual application, due to the narrow field of view of the operating space, the operator relies on the observation method, and there is a large error between the judgment and the actual cutting situation. If the cutting action is ended too early, it is easy to cause cutting failure; if the cutting action is ended too late, the jaws of the ultrasonic knife will continue to wear the gasket, which will cause the ultrasonic knife head to generate a higher temperature and reduce the service life of the ultrasonic knife. Summary of the invention

[0004] The present application provides a shearing point prediction method and ambition based on a fusion model, so that when an ultrasonic knife is shearing an object to be cut, according to the operator's prediction needs, a prediction result of the shearing point moment that meets the needs is given, so as to avoid the situation where the operator fails to observe that the shearing of the object to be cut has been completed, resulting in a higher temperature of the ultrasonic knife head and reducing the service life of the ultrasonic knife.

[0005] The present application provides a shear point prediction method based on a fusion model, comprising:

[0006] Obtaining variable parameters of historical electrical signals of the ultrasonic knife generator in a preset time period, and preprocessing the variable parameters of the historical electrical signals, so that the cutting point of the ultrasonic knife cutting the object to be cut is always within the preset time period;

[0007] The result corresponding to the variable parameter of the historical electrical signal of the first time period is marked as true to obtain a first training set; the result corresponding to the variable parameter of the historical electrical signal of the second time period is marked as true to obtain a second training set, the first time period and the second time period are within the preset time period, and the first time period is earlier than the second time period;

[0008] Constructing a first initial model and a second initial model, and training the first initial model with the first training set to obtain a first model; and training the second initial model with the second training set to obtain a second model;

[0009] When the ultrasonic knife shears the object to be cut, variable parameters of the real-time electrical signal of the ultrasonic knife generator are obtained;

[0010] Using the variable parameters of the real-time electrical signal as inputs of the first model and the second model to obtain a first output of the first model and a second output of the second model;

[0011] The predicted demand is obtained, and the prediction result of the shearing point is obtained according to the predicted demand, the first output and the second output.

[0012] In a possible implementation, the step of obtaining the predicted demand includes:

[0013] Get the thickness of the clipped object;

[0014] The weight corresponding to the predicted demand is determined according to the thickness of the cut object and a preset corresponding relationship, wherein the preset corresponding relationship includes the weight relationship between the thickness of the cut object and the predicted demand.

[0015] In a possible implementation, the following weighted model is used to obtain the prediction result of the shearing point moment, and the weighted model is:

[0016] Y=α″×y1+(1-α″)×y2;

[0017] Wherein, α″ is the weight corresponding to the predicted demand, y1 is the first output, and y2 is the second output.

[0018] In a possible implementation, the step of preprocessing the variable parameters of the historical electrical signal includes:

[0019] Calculate the average value of each parameter in the variable parameters of the historical electrical signal in the third time period, and mark it as the benchmark parameter of each parameter, the third time period is the starting time period in the preset time period;

[0020] Based on the reference parameters of the parameters, each parameter in the variable parameters of the historical electrical signal is adjusted.

[0021] In a possible implementation, the step of preprocessing the variable parameters of the historical electrical signal includes:

[0022] Based on preset rules, outliers in variable parameters of historical electrical signals are removed.

[0023] In a possible implementation, the step of preprocessing the variable parameters of the historical electrical signal includes:

[0024] The exponential average sliding model is used to process the variable parameters of the historical electrical signal. The exponential average sliding model is:

[0025] X t =βX t-1 +(1-β)θ t ;

[0026] Among them, β is the weight parameter, θ t is the weight parameter obtained in the tth update, X t It is the moving average of the variable parameters of the historical electrical signal updated for the tth time.

[0027] In a possible implementation, the step of preprocessing the variable parameters of the historical electrical signal includes:

[0028] The variable parameters of the historical electrical signal are scaled and standardized according to the preset ratio. The standardized processing model is:

[0029]

[0030] Among them, μ is the mean value of each parameter of the variable parameters of the historical electrical signal, and σ is the standard deviation of each parameter of the variable parameters of the historical electrical signal.

[0031] In a possible implementation, the step of preprocessing the variable parameters of the historical electrical signal includes:

[0032] Get the time value of the warning shearing point;

[0033] Based on the timeliness value, the training data in the first training set and the second training set are cut.

[0034] In a possible implementation manner, after constructing the first initial model and the second initial model, the method further includes:

[0035] The first initial model and the second initial model are pre-trained using the variable parameters of the historical electrical signal to obtain basic weights, wherein the batch size of the training input is set to 1024 and the learning rate is set to 0.05.

[0036] In a possible implementation, the step of training the first initial model with the first training set to obtain the first model is specifically as follows:

[0037] Load the basic weights, train the first initial model with the first training set, set the batch size of the training input to 254, set the decay parameter to 0.01, set the learning rate to 0.005, and set the cross entropy loss function to:

[0038] in, is the label value of the real sample in the first training set, p t The predicted probability result of the real shear output by the first initial model is, is the label value of the false sample in the first training set, p fis the predicted probability result of false shearing output by the first initial model, and α is the influencing factor on the label value of the real sample in the first training set.

[0039] In a possible implementation, the step of training the second initial model with the second training set to obtain the second model is specifically as follows:

[0040] Load the base weights, train the second initial model with the second training set, set the batch size of the training input to 254, set the decay parameter to 0.01, set the learning rate to 0.005, and set the cross entropy loss function to:

[0041] in, is the label value of the real sample in the second training set, p t ′ is the predicted probability result of the real shear output by the second initial model, is the label value of the false sample in the second training set, p f ′ is the predicted probability result of false shearing output by the second initial model, and α′ is the impact factor of the real samples in the second training set.

[0042] In one possible implementation, the variable parameters of the electrical signal in the variable parameters of the historical electrical signal and the variable parameters of the real-time electrical signal include: output current value, output voltage value, current-voltage phase difference, the difference between the current-voltage phase difference and the target phase difference, and the first-order derivative of the resonant frequency.

[0043] The present application also provides a shear point prediction device based on a fusion model, the shear point prediction device based on a fusion model is used to execute a shear point prediction method based on a fusion model, and the shear point prediction device based on a fusion model includes:

[0044] The variable parameter acquisition module of the historical electrical signal is used to acquire the variable parameters of the historical electrical signal of the ultrasonic knife generator in a preset time period, and pre-process the variable parameters of the historical electrical signal, so that the cutting point of the ultrasonic knife cutting the object to be cut is always within the preset time period;

[0045] a variable parameter marking module, used to mark the result corresponding to the variable parameter of the historical electrical signal of the first time period as true, so as to obtain a first training set; and used to mark the result corresponding to the variable parameter of the historical electrical signal of the second time period as true, so as to obtain a second training set, the first time period and the second time period are within a preset time period, and the first time period is earlier than the second time period;

[0046] A model building module, used to build a first initial model and a second initial model, and train the first initial model with the first training set to obtain a first model; and used to train the second initial model with the second training set to obtain a second model;

[0047] A variable parameter acquisition module of a real-time electrical signal, which acquires the variable parameters of the real-time electrical signal of the ultrasonic knife generator when the ultrasonic knife is shearing the object to be cut;

[0048] A prediction module, used for taking the variable parameters of the real-time electrical signal as inputs of the first model and the second model to obtain a first output of the first model and a second output of the second model;

[0049] The fusion module is used to obtain the predicted demand and obtain the prediction result of the shearing point according to the predicted demand, the first output and the second output.

[0050] The embodiments of the present application provide a shear point prediction method and device based on a fusion model. Through historical electrical signal parameters, a first model with better prediction advance performance and a second model with better reliability are established. When the ultrasonic knife shears the object to be cut, the variable parameters of the real-time electrical signal of the ultrasonic knife generator are obtained; thereby obtaining the first output of the first model and the second output of the second model; and based on the prediction requirements of the operator, according to the prediction requirements, the first output and the second output, the prediction result of the shear point moment is obtained, and the shear point prediction result that meets the requirements is given, thereby avoiding the situation where the operator fails to observe that the shearing of the object to be cut has been completed, resulting in the ultrasonic knife head generating a higher temperature and reducing the service life of the ultrasonic knife. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solution of the present application, the drawings required for use in the embodiments are briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0052] Figure 1 A schematic diagram of a flow chart of a shear point prediction method based on a fusion model provided in an embodiment of the present application;

[0053] Figure 2 A schematic diagram of a curve showing the resonant frequency and the current-voltage phase difference varying with shear provided in an embodiment of the present application;

[0054] Figure 3 A schematic diagram of the output current and output voltage curves with shearing provided in the embodiment of the present application;

[0055] Figure 4 A schematic diagram of the structure of a fully connected neural network model provided in an embodiment of the present application;

[0056] Figure 5 A schematic diagram of the structure of a recurrent neural network model provided in an embodiment of the present application;

[0057] Figure 6 A schematic diagram of a combined structure of a first model and a second model provided in an embodiment of the present application;

[0058] Figure 7 A schematic diagram of the structure of a shear point prediction device based on a fusion model provided in an embodiment of the present application. DETAILED DESCRIPTION

[0059] When an operator uses an ultrasonic knife to shear an object, in order to accurately predict the time when the object is cut and avoid the operator ending the cutting action too late, which causes the jaws of the ultrasonic knife to continue to wear the gasket, an embodiment of the present application provides a shearing point prediction method and device based on a fusion model.

[0060] The present application embodiment provides a shear point prediction method based on a fusion model, such as Figure 1 As shown, it includes S110 to S160.

[0061] S110, obtaining variable parameters of historical electrical signals of the ultrasonic knife generator in a preset time period, and preprocessing the variable parameters of the historical electrical signals, so that the cutting point when the ultrasonic knife cuts the object to be cut is within the preset time period.

[0062] When the ultrasonic knife is shearing, the electrical signal parameters of the ultrasonic knife generator have obvious characteristics. The output current I is a constant current output, and the output voltage changes with the shearing of the object being cut. Figure 2 As shown in the figure, the phase difference between the resonant frequency and the current and voltage changes with shearing. In order to make the ultrasonic knife work at the highest efficiency, the current phase and voltage phase difference are generally set to 0. In this way, the working circuit of the ultrasonic knife is at the resonant point. As the shearing work proceeds, Figure 3 As shown, the shear point is the output current and output voltage change curve with shearing. The actual impedance changes with the load, and the piezoelectric crystal will also change in real time with the operating temperature.

[0063] After real-time processing by the ultrasonic knife processor, the transducer resonant frequency is adjusted so that the transducer always works in the highest efficiency resonance area. As the shearing is completed, the above voltage and current are collected by the sensor in real time, and the output voltage, output current, resonant frequency, phase difference and other electrical parameters will change slightly. Therefore, these electrical signal parameters can be sent to the processor for further judgment.

[0064] During the shearing process, the output voltage value of the ultrasonic knife changes accordingly with the different materials that the ultrasonic knife head contacts. Specifically, when the ultrasonic knife is shearing the object to be cut, the output voltage value is relatively low. As the shearing progresses, the knife head contacts the Teflon gasket and the output voltage value increases.

[0065] The electrical parameters of current phase value, voltage phase value, target phase value and resonant frequency used in this application are closely related to the changes of the object being cut. Specifically, since the difference between the detected current and voltage phase values ​​and the target phase values ​​is approximately equal to 0 during the stable shearing of the object being cut by the ultrasonic knife, the rate of change of the resonant frequency is relatively constant. As the shearing progresses, the blade contacts the Teflon gasket, the difference between the current and voltage phase values ​​and the target phase values ​​will deviate from the value of 0, the rate of change of the resonant frequency becomes larger, and the waveform produces larger fluctuations. In this way, using the resonant frequency as the collected electrical signal parameter has a more obvious distinguishing feature.

[0066] Among them, the historical electrical signal refers to the output current value, output voltage value, current phase value, voltage phase value, target phase value and resonant frequency; in order to obtain training data with higher information content and better feature expression, the historical electrical signal can be further processed to obtain the variable parameters of the historical electrical signal. For example, the current phase value and the voltage phase value are processed to obtain the current-voltage phase difference; the current-voltage phase difference and the target phase value are processed to obtain the difference between the current-voltage phase difference and the target phase difference; the first-order derivative of the resonant frequency is calculated. In this way, the variable parameters of the historical electrical signal are: output current value, output voltage value, current-voltage phase difference, the difference between the current-voltage phase difference and the target phase difference, and the first-order derivative of the resonant frequency.

[0067] It should be noted that the present application collects the current phase value, voltage phase value, target phase value, and resonant frequency of the ultrasonic knife generator because the above-mentioned electrical signal parameters are closely related to the changes in the object being cut. In actual application, it is not limited to collecting only the above-mentioned electrical signal parameters. For example, in view of the differences in the circuit properties of the ultrasonic knife generator, if there are other electrical signal parameters that are closely related to the changes in the object being cut, they can also be used as the collected electrical signal parameters. For another example, the collected electrical signal parameters can also include multi-dimensional variables such as sensor collected temperature, resistance, ultrasonic knife parameters, or generator parameters.

[0068] In addition, when determining the variable parameters of the historical electrical signal, the ratio of voltage to current can also be calculated to obtain the impedance value, and the product of voltage and current can be calculated to obtain the power value, and the impedance value and power value can be used as the variable parameters of the historical electrical signal.

[0069] In the subsequent process of training the first initial model and the second initial model, in order to enable the first initial model and the second initial model to better learn the deep features of the training data, the present application scales the input training data to scale the variable parameters of the historical electrical signal to a similar scale.

[0070] For example, when the resonant frequency is around 55000 Hz and the current-voltage phase difference is within the range of ±180°, if the first-order derivative of the resonant frequency and the current-voltage phase difference are directly input into the first initial model and the second initial model for training, it is difficult for the first initial model and the second initial model to learn effective features.

[0071] In this regard, the present embodiment uses a scaling method to process the historical electrical signal or the variable parameters of the historical electrical signal, for example, first subtract 55000 Hz from the resonant frequency, and then multiply it by a reduction factor of 0.1 to obtain a new frequency variable. In this way, the new frequency variable has the same order of magnitude as the phase difference variable.

[0072] In addition, during the data collection process, the collected electrical signal parameters of the ultrasonic knife generator have a certain offset. In order to avoid the influence of the offset on the data characteristics. In some embodiments, first, the standard mean and standard deviation of each parameter in the historical electrical signal are obtained, and then the standard mean and standard deviation are used to standardize each parameter in the historical electrical signal. The corresponding standardization model is as follows:

[0073]

[0074] Among them, the standardized processing model is: x is each parameter in the historical electrical signal, μ is the mean value of each parameter in the historical electrical signal, and σ is the standard deviation of each parameter in the historical electrical signal.

[0075] The variable parameters of the historical electrical signals obtained in the embodiment of the present application are used for model training. In the actual application process, a certain amount of data will be obtained. For example, in the process of multiple shearing with an ultrasonic knife, relevant parameters are collected. The preset time needs to include the time when the cut object is cut (the moment of the shearing point). For example, the time interval for data collection is 10ms, and the data collected in the preset time is 400 data points before the shearing point to 50 data points after the shearing point.

[0076] In an embodiment of the present application, when collecting the electrical signal of the ultrasonic knife generator, it is accompanied by a frequency locking process. Therefore, in the initial stage of collection, the collected electrical signal of the ultrasonic knife generator is unstable. In order to solve this problem, the present application includes the following steps in the step of preprocessing the variable parameters of the historical electrical signal: calculating the average value of each parameter in the variable parameters of the historical electrical signal in the third time period, and marking it as the benchmark parameter of each parameter, and the third time period is the starting time period in the preset time period; based on the benchmark parameters of each parameter, adjusting each parameter in the variable parameters of the historical electrical signal. For example, the average value of the parameters collected in the first 0.5 seconds is calculated, and the average value of the parameters collected in the first 0.5 seconds is set as the benchmark, and the parameters collected after 0.5 seconds minus the benchmark are used as valid data.

[0077] Since the electrical signal of the ultrasonic knife generator is easily affected by various uncertain factors such as temperature and load, the waveform of the electrical signal is not stable. In this regard, in some embodiments, the step of preprocessing the variable parameters of the historical electrical signal includes: using a second exponential average sliding model to process the variable parameters of the historical electrical signal, and the corresponding exponential average sliding model is:

[0078] X t =βX t-1 +(1-β)θ t .

[0079] Among them, β is the weight parameter, θ t is the weight parameter of each parameter in the historical electrical signal updated at the tth time, X t is the moving average of the variable parameters of the historical electrical signal updated for the tth time. For example, the signal is smoothed by using an average sliding window with a window size of 30. After processing, the signal does not change the original trend, but is smoother after removing high-frequency small fluctuations. In addition, the variable parameters of the historical electrical signal can also be processed by moving average method, low-pass filtering method, polynomial fitting method, local weighted scatter point smoothing method or Kalman filtering method.

[0080] S120, marking the result corresponding to the variable parameter of the historical electrical signal of the first time period as true to obtain a first training set; marking the result corresponding to the variable parameter of the historical electrical signal of the second time period as true to obtain a second training set, the first time period and the second time period are within the preset time period, and the first time period is earlier than the second time period.

[0081] In the embodiment of the present application, the data collection time of the first training set is earlier, for example, the variable parameters of the historical electrical signal in the first time period are the data points 200 before the cutting point to the data points 100 before the cutting point. In this way, the model trained with the first training set has a better prediction performance in advance. The data collection time of the second training set is later, for example, the variable parameters of the historical electrical signal in the second time period are the data points 50 before the cutting point to the data points 50 after the cutting point. In this way, the model trained with the second training set has a more reliable prediction performance.

[0082] It should be noted that after marking the result corresponding to the variable parameters of some historical electrical signals as true, it is also necessary to mark the result corresponding to the variable parameters of the historical electrical signals in other stages as false. Among them, the result corresponding to the variable parameters of the historical electrical signals is true, which means that the variable parameters of the historical electrical signals correspond to the cut object being cut, and the result corresponding to the variable parameters of the historical electrical signals is false, which means that the variable parameters of the historical electrical signals correspond to the cut object being not cut. It should be noted that after marking the result corresponding to the variable parameters as true, the corresponding variable parameters are real samples, and after marking the result corresponding to the variable parameters as false, the corresponding variable parameters are false samples.

[0083] In the embodiment of the present application, it is necessary to cut a large amount of real and effective data collected, and use the cut data to train the model. In order to take advantage of the training equipment that can be trained in batches, it is necessary to make the training input data have a uniform length. In the step of preprocessing the variable parameters of the historical electrical signal, it includes: obtaining the timeliness value at the warning cutting point; based on the timeliness value, cutting the training data in the first training set and the second training set.

[0084] For example, based on the characteristics of the object being cut, the timeliness of the shearing warning is required to be less than 1 second. At the same time, the electrical signal of the ultrasonic knife generator has a strong correlation in time. This application sets the shearing time of the training data of the first training set and the second training set to 1 second, that is, the amount of data with a shearing time of 1 second is set as the uniform length of the training data. And the length of the training data is 2 n When the device used to train the model has better parallelism, the present application adopts 128 as the uniform length of the training input data. On the other hand, the present invention slides and intercepts a segment of length 128 in the original data with a step size of 16 as the training input of the subsequent first initial model and the second initial model, which can effectively increase the number of data samples in the training set and enable the fusion model to fully learn the deep data features.

[0085] S130, constructing a first initial model and a second initial model, and training the first initial model with the first training set to obtain a first model; and training the second initial model with the second training set to obtain a second model.

[0086] Among them, the first model and the second model involved in this application are two models with similar structures and different weights, in order to obtain the first model and the second model with different weights. After constructing the first initial model and the second initial model, first, the first initial model and the second initial model are pre-trained using the variable parameters of the historical electrical signal to obtain the basic weights, wherein during the pre-training process, the batch size of the training input is set to 1024, and the learning rate is set to 0.05. Then, the basic weights are loaded, and the first initial model after loading the basic weights is trained with the training data in the first training set. Among them, in the process of training the first initial model after loading the basic weights, the batch size of the training input is set to 254, the attenuation parameter is set to 0.01, the learning rate is set to 0.005, and the cross entropy loss function is set to: in, is the label value of the real sample in the first training set, p t The predicted probability result of the real shear output by the first initial model is, is the label value of the false sample in the first training set, p f is the predicted probability result of false shearing output by the first initial model, and α is the influencing factor on the label value of the real sample in the first training set.

[0087] And, load the basic weights, and train the second initial model after loading the basic weights with the training data in the second training set, wherein, in the process of training the second initial model after loading the basic weights, set the batch size of the training input to 254, set the decay parameter to 0.01, set the learning rate to 0.005, and set the cross entropy loss function to: in, is the label value of the real sample in the second training set, p t ′ is the predicted probability result of the real shear output by the second initial model, is the label value of the false sample in the second training set, p f ′ is the predicted probability result of false shearing output by the second initial model, and α′ is the impact factor of the real samples in the second training set.

[0088] Among them, the use of the cross entropy loss function with an influencing factor can increase the loss of the label value of the real sample, affect the model's learning of the real sample, improve the overall accuracy, and avoid the situation where the real sample occupies a very small part of the training data, resulting in unbalanced training labels.

[0089] Among them, in the embodiment of the present application, when α=1.2 and α′=1.2, the first initial model and the second initial model have better training effects. In addition, the loss function is not limited to the cross entropy loss function, and the mean square error loss function, the square loss function or the absolute value loss function may also be used.

[0090] S140, when the ultrasonic knife shears the object to be cut, obtaining variable parameters of the real-time electrical signal of the ultrasonic knife generator.

[0091] It should be noted that before the variable parameters of the real-time electrical signal are input into the first model and the second model, it is necessary to perform necessary preprocessing on the variable parameters of the real-time electrical signal, and the preprocessing process is the same as the variable parameter processing process of the historical electrical signal when training the first initial model and the second initial model, that is, referring to the above-mentioned preprocessing process for the variable parameters of the historical electrical signal, before the variable parameters of the real-time electrical signal are input into the first model or the second model, the same preprocessing process is performed on the variable parameters of the real-time electrical signal.

[0092] S150, using the variable parameters of the real-time electrical signal as inputs of the first model and the second model to obtain a first output of the first model and a second output of the second model.

[0093] Among them, the first output and the second output are probability values ​​that characterize whether the prediction result of the cutting point of the cut object is true, and the probability value that characterizes the true value can be a probability value ranging from 0 to 1 or a probability value set with a sum of 1.

[0094] The first output of the first model has better prediction advance performance, that is, it can give the prediction result of the shearing point earlier before the shearing point occurs. The second output of the second model has higher reliability, that is, the shearing point result given is more reliable.

[0095] S160, obtaining predicted demand, and obtaining a predicted result of the shearing point according to the predicted demand, the first output and the second output.

[0096] Among them, the prediction demand mainly reflects the operator's expectations for the prediction results. For example, the operator's prediction demand is to know the prediction results of the shearing point earlier, so as to control the parameter adjustment of the ultrasonic knife in advance. For another example, the operator's demand is that the shearing point prediction results are more reliable to ensure more accurate operation of the ultrasonic knife.

[0097] In actual application, after using the variable parameters of the real-time electrical signal as the input of the first model and the second model to obtain the first output of the first model and the second output of the second model, the result of the first output can be independently judged. For example, if the signal feature recognition of the first output characterizing shearing is successful, the prediction result of the shearing point is given, and the operator is directly prompted by sound, light, electricity, etc., so that the operator can implement protective measures such as limiting the output power of the ultrasonic knife generator.

[0098] Alternatively, after the signal feature of the first output representing shearing is successfully identified, the result of the second output is judged. If the signal feature of the second output representing shearing is successfully identified, the prediction result of the shearing point is given in a weighted manner according to the prediction requirements.

[0099] Among them, the weighted model is: Y=α″×y1+(1-α″)×y2, wherein α″ is the weight corresponding to the prediction demand, y1 is the first output, y2 is the second output, and the weight α″ is a set of values ​​that can be adjusted according to the application scenario. The operator can determine the weight value that reflects the prediction demand according to the characteristics of the object to be cut. For example, setting α″=0.9 makes the judgment result more advanced when dealing with thicker objects to be cut. For another example, setting α″=0.7 makes the judgment result more reliable when dealing with thinner objects to be cut.

[0100] In other words, the weight of the predicted demand can be determined by the thickness of the cut object. For example, first, the thickness of the cut object is obtained, and then the weight corresponding to the predicted demand is determined according to the thickness of the cut object and the preset correspondence, that is, the value of the predicted demand α″ is determined, wherein the preset correspondence includes the weight relationship between the thickness of the cut object and the predicted demand.

[0101] An embodiment of the present application provides a shear point prediction method based on a fusion model. Through historical electrical signal parameters, a first model with better prediction advance performance and a second model with better reliability are established. When the ultrasonic knife shears the object to be cut, the variable parameters of the real-time electrical signal of the ultrasonic knife generator are obtained; thereby obtaining the first output of the first model and the second output of the second model; and based on the prediction needs of the operator, according to the prediction needs, the first output and the second output, the prediction result of the shear point moment is obtained, and the shear point prediction result that meets the needs is given, thereby avoiding the situation where the operator fails to observe that the shearing of the object to be cut has been completed, resulting in a higher temperature of the ultrasonic knife head and reducing the service life of the ultrasonic knife.

[0102] In the fusion model-based cutting point prediction method provided in the embodiment of the present application, the first model or the second model may be a combination of one or more of a machine learning algorithm model or a deep learning algorithm model. More specifically, the machine learning algorithm may include a linear regression algorithm, a support vector machine algorithm, a nearest neighbor algorithm, a decision tree algorithm, a random forest algorithm, or a naive Bayes algorithm, and the deep learning algorithm may include a fully connected neural network, a convolutional neural network, a generative adversarial network, or a recurrent neural network.

[0103] For example, the first model adopts a combination of a fully connected neural network model formed by a fully connected neural network within the above optional range and a recurrent neural network model formed by a recurrent neural network.

[0104] Among them, the fully connected neural network model is a mathematical model inspired by neurons, which is composed of multiple nodes connected to each other and can be used to model complex relationships between data. The fully connected neural network model has a multi-layer structure, and all neurons in each layer are connected to each other. For example, Figure 4 As shown, the fully connected neural network model used in the embodiment of the present application includes an input layer, a hidden layer and an output layer.

[0105] Furthermore, the fitting ability of the fully connected neural network model is highly correlated with the number of layers and neurons. The more layers there are and the richer the number of neurons, the stronger the mathematical fitting ability of the fully connected neural network model. Figure 4 The two hidden layers shown, each hidden unit corresponds to the number of output units of the recurrent neural network. Generally, the number of neurons is set to 8, 16, 32, 64 or 128. In the embodiment of the present application, the number of neurons is set to 16.

[0106] Among them, considering that the collected ultrasonic knife electrical signal is a signal with strong time sequence, in order to better process the information relationship in time sequence, in some embodiments, such as Figure 5 As shown, a recurrent neural network model with memory capability can also be used to construct the first initial model or the second initial model. The recurrent neural network model can memorize the previous signal and has a stronger feature extraction capability for the time series signal. In this way, the first initial model or the second initial model is constructed using the recurrent neural network model and the fully connected neural network model.

[0107] For example, considering that the electrical signal of the ultrasonic generator in practical applications has temporal continuity, since the electrical signal of the ultrasonic generator has a stronger correlation in a short time, in some embodiments, the time interval for collecting the electrical signal of the ultrasonic generator is set to 10 milliseconds, so the number of loop nodes of the designed recurrent neural network is 40. In this way, the electrical signals within the time interval of 400 milliseconds before and after can be processed simultaneously, and the correlation of the output results on the time scale can be strengthened.

[0108] It should be noted that the first model and the second model of the present application can be various combinations of machine learning algorithm models and deep learning algorithm models, or combinations of the same type of algorithm models. The combination method is parallel, series or cross. Figure 6 As shown, it is a specific method adopted by the embodiment of the present application, which is a model structure of the first model and the second model constructed above in parallel. In particular, the above two basic model structures are the same, but the training data are different, the weight parameters are different, and the discrimination results obtained by forward reasoning are different.

[0109] The embodiment of the present application also provides a shear point prediction device based on a fusion model, and the shear point prediction device based on a fusion model is used to execute the shear point prediction method based on the fusion model, such as Figure 7 As shown, the shear point prediction device types based on the fusion model include:

[0110] The variable parameter acquisition module of the historical electrical signal is used to acquire the variable parameters of the historical electrical signal of the ultrasonic knife generator in a preset time period, and pre-process the variable parameters of the historical electrical signal, so that the cutting point of the ultrasonic knife cutting the object to be cut is always within the preset time period;

[0111] a variable parameter marking module, used to mark the result corresponding to the variable parameter of the historical electrical signal of the first time period as true, so as to obtain a first training set; and used to mark the result corresponding to the variable parameter of the historical electrical signal of the second time period as true, so as to obtain a second training set, the first time period and the second time period are within a preset time period, and the first time period is earlier than the second time period;

[0112] A model building module, used to build a first initial model and a second initial model, and train the first initial model with the first training set to obtain a first model; and used to train the second initial model with the second training set to obtain a second model;

[0113] A variable parameter acquisition module of a real-time electrical signal, which acquires the variable parameters of the real-time electrical signal of the ultrasonic knife generator when the ultrasonic knife is shearing the object to be cut;

[0114] A prediction module, used for taking the variable parameters of the real-time electrical signal as inputs of the first model and the second model to obtain a first output of the first model and a second output of the second model;

[0115] The fusion module is used to obtain the predicted demand and obtain the prediction result of the shearing point according to the predicted demand, the first output and the second output.

[0116] The embodiments of the present application provide a shear point prediction method and device based on a fusion model. Through historical electrical signal parameters, a first model with better prediction advance performance and a second model with better reliability are established. When the ultrasonic knife shears the object to be cut, the variable parameters of the real-time electrical signal of the ultrasonic knife generator are obtained; thereby obtaining the first output of the first model and the second output of the second model; and based on the prediction requirements of the operator, according to the prediction requirements, the first output and the second output, the prediction result of the shear point moment is obtained, and the shear point prediction result that meets the requirements is given, thereby avoiding the situation where the operator fails to observe that the shearing of the object to be cut has been completed, resulting in the ultrasonic knife head generating a higher temperature and reducing the service life of the ultrasonic knife.

[0117] The above specific implementation methods further illustrate the purpose, technical solutions and beneficial effects of the present application in detail. It should be understood that the above are only specific implementation methods of the present application and are not used to limit the protection scope of the present application. Any modifications, equivalent substitutions, improvements, etc. made on the basis of the technical solutions of the present application should be included in the protection scope of the present application.

Claims

1. A shear point prediction method based on a fusion model, characterized in that: include: Obtaining variable parameters of historical electrical signals of the ultrasonic knife generator in a preset time period, and preprocessing the variable parameters of the historical electrical signals, wherein the cutting point of the ultrasonic knife cutting the object to be cut is within the preset time period; The result corresponding to the variable parameter of the historical electrical signal in the first time period is marked as true to obtain a first training set; the result corresponding to the variable parameter of the historical electrical signal in the second time period is marked as true to obtain a second training set, the first time period and the second time period are within a preset time period, and the first time period is earlier than the second time period; Constructing a first initial model and a second initial model, and training the first initial model with a first training set to obtain a first model; and training the second initial model with a second training set to obtain a second model; When the ultrasonic knife shears the object to be cut, obtaining variable parameters of the real-time electrical signal of the ultrasonic knife generator; Using the variable parameters of the real-time electrical signal as inputs of the first model and the second model to obtain a first output of the first model and a second output of the second model; A predicted demand is obtained, and a predicted result of a shearing point is obtained according to the predicted demand, the first output, and the second output.

2. The shear point prediction method based on the fusion model according to claim 1 is characterized in that: The step of obtaining the predicted demand includes: Get the thickness of the clipped object; The weight corresponding to the prediction demand is determined according to the thickness of the cut object and a preset corresponding relationship, wherein the preset corresponding relationship includes a weight relationship between the thickness of the cut object and the prediction demand.

3. The shear point prediction method based on the fusion model according to claim 2 is characterized in that: The following weighted model is used to obtain the prediction result of the shearing point moment. The weighted model is: Y=α″×y1+(1-α″)×y2; Wherein, α″ is the weight corresponding to the predicted demand, y1 is the first output, and y2 is the second output.

4. The shear point prediction method based on the fusion model according to claim 1 is characterized in that: The step of preprocessing the variable parameters of the historical electrical signal includes: Calculate the average value of each parameter in the variable parameters of the historical electrical signal in the third time period, and mark it as the reference parameter of each parameter, wherein the third time period is the starting time period in the preset time period; Based on the reference parameters of the parameters, each parameter in the variable parameters of the historical electrical signal is adjusted.

5. The shear point prediction method based on the fusion model according to claim 1 is characterized in that: The step of preprocessing the variable parameters of the historical electrical signal includes: Based on preset rules, outliers in the variable parameters of the historical electrical signal are removed.

6. The shear point prediction method based on the fusion model according to claim 1 is characterized in that: The step of preprocessing the variable parameters of the historical electrical signal includes: An exponential average sliding model is used to process the variable parameters of the historical electrical signal, and the exponential average sliding model is: X t =βX t-1 +(1-β)θ t ; Among them, β is the weight parameter, θ t is the weight parameter obtained in the tth update, X t It is the moving average of the variable parameters of the historical electrical signal updated for the tth time.

7. The shear point prediction method based on the fusion model according to claim 1 is characterized in that: The step of preprocessing the variable parameters of the historical electrical signal includes: Get the timeliness value of the warning shearing point; Based on the timeliness value, the training data in the first training set and the second training set are cut.

8. The shear point prediction method based on the fusion model according to claim 1 is characterized in that: After constructing the first initial model and the second initial model, it also includes: The first initial model and the second initial model are pre-trained using the variable parameters of the historical electrical signal to obtain basic weights, wherein the batch size of the training input is set to 1024 and the learning rate is set to 0.

05.

9. The shear point prediction method based on the fusion model according to claim 8 is characterized in that: The step of training the first initial model with the first training set to obtain the first model is specifically as follows: Load the basic weights, train the first initial model with the first training set, set the batch size of the training input to 254, set the decay parameter to 0.01, set the learning rate to 0.005, and set the cross entropy loss function to: in, is the label value of the real sample in the first training set, p t The predicted probability result of the real shear output by the first initial model is, is the label value of the false sample in the first training set, p r is the predicted probability result of false shearing output by the first initial model, and α is the influencing factor of the label value of the real sample in the first training set; The step of training the second initial model with the second training set to obtain the second model is specifically: Load the basic weights, train the second initial model with the second training set, set the batch size of the training input to 254, set the decay parameter to 0.01, set the learning rate to 0.005, and set the cross entropy loss function to: in, is the label value of the real sample in the second training set, pt′ is the predicted probability result of the real cut output by the second initial model, is the label value of the false sample in the second training set, pr′ is the predicted probability result of false shearing output by the second initial model, and α′ is the impact factor on the real samples in the second training set.

10. The shear point prediction method based on fusion model according to claim 1, characterized in that: The variable parameters of the electrical signal in the variable parameters of the historical electrical signal and the variable parameters of the real-time electrical signal include: output current value, output voltage value, current-voltage phase difference, the difference between the current-voltage phase difference and the target phase difference, and the first-order derivative of the resonant frequency.

11. A shear point prediction device based on a fusion model, characterized in that: The shear point prediction device based on the fusion model is used to execute the shear point prediction method based on the fusion model according to any one of claims 1 to 10, and the shear point prediction device based on the fusion model includes: A variable parameter acquisition module of a historical electrical signal, used to acquire the variable parameters of the historical electrical signal of the ultrasonic knife generator in a preset time period, and pre-process the variable parameters of the historical electrical signal, wherein the cutting point of the ultrasonic knife cutting the object to be cut is within the preset time period; a variable parameter marking module, used to mark the result corresponding to the variable parameter of the historical electrical signal in the first time period as true, so as to obtain a first training set; and used to mark the result corresponding to the variable parameter of the historical electrical signal in the second time period as true, so as to obtain a second training set, wherein the first time period and the second time period are within a preset time period, and the first time period is earlier than the second time period; A model building module, configured to build a first initial model and a second initial model, and train the first initial model with a first training set to obtain a first model; and to train the second initial model with a second training set to obtain a second model; A variable parameter acquisition module of a real-time electrical signal, which acquires the variable parameters of the real-time electrical signal of the ultrasonic knife generator when the ultrasonic knife is shearing the object to be cut; A prediction module, used for taking the variable parameters of the real-time electrical signal as inputs of the first model and the second model to obtain a first output of the first model and a second output of the second model; The fusion module is used to obtain the predicted demand and obtain the prediction result of the shearing point according to the predicted demand, the first output and the second output.

Images