A machine tool fault prediction and diagnosis method based on deep learning

Abstract

The present invention discloses a machine tool fault prediction and diagnosis method based on deep learning, including: collecting machine tool data to obtain pulse factor, volatility, spectrum energy and spectrum bandwidth; generating data features of a target machine tool according to the pulse factor, volatility, spectrum energy and spectrum bandwidth; performing linear transformation on the data features according to a pre-built deep learning model to obtain primary feature output; performing secondary transformation on the primary feature output to obtain dual transformation features of the primary feature output; performing parameter update on the pre-built deep learning model according to the data features and the dual transformation features to obtain an updated deep learning model; collecting real-time data of the target machine tool, inputting the real-time data into the updated deep learning model, obtaining the fault probability of the target machine tool, and maintaining the target machine tool according to the fault probability. The present invention can improve the efficiency of machine tool fault prediction and diagnosis.

Description

Technical Field

[0001] The present invention belongs to the field of machine tool fault detection, and in particular relates to a machine tool fault prediction and diagnosis method based on deep learning. Background Art

[0002] In modern manufacturing, machine tool equipment is one of the core links of production. Its failure will cause equipment downtime, which in turn affects the operation of the entire production line. Traditional fault diagnosis methods rely on manual inspections and regular maintenance. This passive maintenance mode not only reduces production efficiency, but also increases maintenance costs. By introducing efficient fault prediction and diagnosis technology, "predictive maintenance" of equipment can be achieved, that is, early warning before the failure occurs, maintenance in advance, and avoiding production stagnation.

[0003] For example, the Chinese patent document with publication number CN117407814A discloses a machine tool equipment fault warning method and system based on digital twins; the Chinese patent document with publication number CN118277884A discloses an intelligent diagnosis method for CNC machine tool faults based on graph theory algorithm and ACO-SVM.

[0004] However, many existing machine tool fault diagnosis systems rely on historical data during training, and in actual applications, factors such as the working environment of the equipment, differences in operators, and changes in material properties may lead to insufficient generalization of the model in new environments. Especially in the case of small samples and frequent abnormal data, traditional algorithms may have overfitting or underfitting problems, thus affecting the real-time performance of fault diagnosis and prediction. Therefore, how to improve the efficiency of machine tool fault prediction and diagnosis has become an urgent problem to be solved. Summary of the invention

[0005] The present invention provides a machine tool fault prediction and diagnosis method based on deep learning, which can solve the problem of low efficiency in machine tool fault prediction and diagnosis.

[0006] A machine tool fault prediction and diagnosis method based on deep learning, comprising the following steps:

[0007] (1) Collect machine tool data to obtain pulse factor, fluctuation rate, spectrum energy and spectrum bandwidth;

[0008] (2) Generate data features of the target machine tool based on pulse factor, fluctuation rate, spectrum energy, and spectrum bandwidth;

[0009] (3) Perform linear transformation on data features according to the pre-built deep learning model to obtain primary feature output;

[0010] (4) Perform secondary transformation on the primary feature output to obtain double transformed features of the primary feature output;

[0011] (5) Update the parameters of the pre-built deep learning model according to the data features and the dual transformation features to obtain an updated deep learning model;

[0012] (6) Collect the real-time data of the target machine tool and input the real-time data into the updated deep learning model to obtain the failure probability of the target machine tool.

[0013] In step (1), the calculation formula of the pulse factor is as follows:

[0014] ;

[0015] in, is the pulse factor of the target machine tool, It is the machine data. is the maximum number in the pre-collected machine data sequence, is the standard deviation of the pre-collected machine data series.

[0016] The volatility calculation formula is as follows:

[0017] ;

[0018] in, is the volatility of the target machine tool, It is the machine data. is the maximum number in the pre-collected machine data sequence, is the minimum number in the pre-collected machine data sequence.

[0019] The calculation formula of spectrum energy is as follows:

[0020] ;

[0021] in, is the spectral energy of the target machine tool, is the total number of machine data. is the frequency domain signal amplitude of the machine tool data, is the discrete frequency point of the machine tool data, It is the frequency point index of the machine tool data.

[0022] The spectrum bandwidth is calculated as follows:

[0023] ;

[0024] in, is the spectrum bandwidth of the target machine tool, is the center frequency of the machine data, is the total number of machine data. is the frequency domain signal amplitude of the machine tool data, is the discrete frequency point of the machine tool data, It is the frequency point index of the machine tool data.

[0025] In step (2), after the pulse factor, fluctuation rate, spectrum energy and spectrum bandwidth are fused, the data features of the target machine tool are obtained.

[0026] In step (3), the data features are linearly transformed according to the linear transformation algorithm in the pre-built deep learning model to obtain the primary feature output of the data features, wherein the linear transformation algorithm is:

[0027] ;

[0028] in, is the primary feature output of the data feature, o is the identifier of the data feature, is the hidden layer neuron index in the pre-built deep learning model, is the total number of data features, are the weights from the input layer to the hidden layer in the pre-built deep learning model, is the bias of the hidden layer in the pre-built deep learning model, is the data feature, is the activation function.

[0029] In step (4), the primary feature output is secondary transformed using a preset quadratic linear transformation algorithm to obtain a double transformation feature of the primary feature output, wherein the preset quadratic linear transformation algorithm is:

[0030] ;

[0031] in, is the double transformation feature of the primary feature output, is the hidden layer neuron index in the pre-built deep learning model, is the total number of hidden layer neurons in the pre-built deep learning model, is the primary feature output, are the weights from the hidden layer to the output layer in the pre-built deep learning model, is the bias of the output layer.

[0032] The specific process of step (5) is as follows:

[0033] (5-1) Calculating the parameter gradient of the preset mean square error loss function with respect to the model parameters of the deep learning model according to the data features and the dual transformation features;

[0034] (5-2) The parameters of the pre-built deep learning model are updated according to the parameter gradient and the preset update algorithm to obtain an updated deep learning model.

[0035] In step (5-2), the preset update algorithm is:

[0036] ;

[0037] in, It is The model parameters of the iteration, It is The model parameters of the iteration, is the learning rate, is the number of iterations, It is The parameter gradient of the preset mean square error loss function with respect to the model parameters of the deep learning model after iterations.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. The present invention uses sophisticated feature extraction, such as pulse factor, volatility, spectral energy and spectral bandwidth, which can accurately reflect changes in the operating status of the machine tool and capture potential fault signals. This process not only relies on traditional statistical analysis methods, but also uses deep learning models to perform multi-dimensional linear and nonlinear transformations on data, so that the model can identify more complex fault modes and potential risks.

[0040] 2. In the present invention, the deep learning model can automatically extract important patterns and rules from historical operation data through adaptive learning and continuous optimization mechanisms, avoiding the limitations of artificial feature design and enhancing the model's prediction and generalization capabilities. In addition, with the input of real-time data, the model can perform online learning and parameter updates during operation, and continuously adjust to adapt to new working environments and changes in fault types. This dynamic adjustment capability is unattainable by traditional methods. Through this online learning, the deep learning model can adjust the prediction strategy based on real-time data.

[0041] 3. In the present invention, the fault probability value output by the deep learning model can quantify the possibility and occurrence time of machine tool failure, providing a scientific basis for subsequent maintenance decisions, thereby avoiding the problems of misjudgment and missed judgment in traditional diagnostic methods and improving the accuracy and reliability of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a flow chart of a machine tool fault prediction and diagnosis method based on deep learning in the present invention. DETAILED DESCRIPTION

[0043] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be pointed out that the embodiments described below are intended to facilitate the understanding of the present invention and do not have any limiting effect on the present invention.

[0044] like Figure 1 As shown, a machine tool fault prediction and diagnosis method based on deep learning includes the following steps:

[0045] S1. Generate the pulse factor and fluctuation rate of the pre-collected machine tool data, and generate the spectrum energy and spectrum bandwidth of the machine tool data.

[0046] In the embodiment of the present invention, the pulse factor of the target machine tool is generated by the algorithm:

[0047] ;

[0048] in, is the pulse factor of the target machine tool, is the maximum number in the pre-collected machine data sequence, is the standard deviation of the pre-collected machine data sequence, It is machine tool data.

[0049] In detail, the pulse factor is an indicator used to measure the pulse characteristics in machine tool data. It is obtained by calculating the ratio of the maximum value in the data sequence to the standard deviation. This indicator can reflect whether there is an instantaneous strong impact or mutation during the operation of the machine tool. For example, when a tool is suddenly damaged or there is an abnormal collision between components, a large pulse factor value will often be reflected in the corresponding monitoring data (such as vibration data, current data, etc.), thereby helping to determine whether the machine tool is in an abnormal state and predict possible faults.

[0050] In detail, It refers to the maximum value in the pre-collected machine tool data sequence. The machine tool data here can be time series data collected from different sensors, such as data on the vibration amplitude changing over time collected by a vibration sensor. The maximum amplitude found in this set of vibration amplitude data is what is mentioned here. It reflects the extreme value of the data in a certain dimension, which is critical for capturing abnormal peaks in the data.

[0051] In detail, It is the standard deviation of the pre-collected machine tool data sequence, which is a statistic used to measure the degree of dispersion or fluctuation of machine tool data. The standard deviation reflects the degree of dispersion of the data relative to the mean. The larger the standard deviation, the more drastic the fluctuation of the data. In the calculation of the pulse factor, considering it together with the maximum value can more comprehensively analyze the relationship between the pulse characteristics and the overall fluctuation in the data.

[0052] Specifically, machine tool data refers to the raw data reflecting the operating status of the machine tool collected by installing sensors at various key parts of the machine tool. Common data include vibration acceleration, velocity or displacement data collected by vibration sensors, temperature data of various components collected by temperature sensors, motor operating current data collected by current sensors, etc. These data are recorded in the form of time series to provide basic information for subsequent fault prediction and diagnosis.

[0053] In the embodiment of the present invention, the fluctuation rate of the target machine tool is generated by the following algorithm:

[0054] ;

[0055] in, is the volatility of the target machine tool, is the maximum number in the pre-collected machine data sequence, is the minimum number in the pre-collected machine data sequence, It is machine tool data.

[0056] Specifically, volatility is an indicator that simply and intuitively describes the fluctuation range of machine tool data, which is obtained by calculating the difference between the maximum and minimum values ​​in the data sequence. The larger the range, the greater the fluctuation of the data in a given time period, indicating that the operating state of the machine tool is more unstable, which may indicate that the machine tool has problems such as worn or loose parts, or is subject to external interference, which helps to make preliminary judgments on potential machine tool failures.

[0057] In detail, It is the minimum number in the pre-collected machine tool data sequence, reflecting the lower limit of the data in the time series, and together with the maximum value, it determines the fluctuation range of the data, that is, the size of the range (volatility).

[0058] In detail, by calculating these pulse factors and fluctuation rates, the key features of the machine tool data can be extracted from the time domain perspective, and then combined with the subsequent generated frequency domain features such as spectral energy and spectral bandwidth, strong data support can be provided for more comprehensive and accurate prediction and diagnosis of machine tool faults.

[0059] In the embodiment of the present invention, the spectrum energy is generated by the following algorithm:

[0060] ;

[0061] in, is the spectral energy of the target machine tool, is the total number of machine data. is the frequency domain signal amplitude of the machine tool data, is the discrete frequency point of the machine tool data, It is the frequency point index of the machine tool data.

[0062] Specifically, spectrum energy is an indicator used to measure the energy distribution of machine tool data in the frequency domain. After the data collected by the machine tool is transformed in the frequency domain (such as Fourier transform, etc.), a value is calculated according to a specific algorithm. Its size reflects the total energy of the information related to the machine tool operation status in the entire frequency domain. Different failure modes often cause characteristic changes in spectrum energy in different frequency bands. For example, when a component has a wear failure, the energy at a specific frequency component may increase or decrease significantly, so spectrum energy can be used as an important basis for subsequent fault diagnosis.

[0063] In detail, when performing frequency domain analysis, the continuous frequency range is usually discretized to obtain a finite number of discrete frequency points for calculation. It refers to the number of these discrete frequency points. For example, after performing a discrete Fourier transform (DFT) on the vibration signal of a machine tool for a period of time, we get The spectrum information corresponding to the frequency points covers the entire frequency range of interest. The value of will affect the resolution and accuracy of frequency domain analysis. It is generally determined based on actual needs and relevant principles of signal processing. Common ones are ( is a positive integer), such as .

[0064] In detail, the frequency domain signal amplitude is expressed at a frequency of The frequency domain signal amplitude at the frequency domain, that is, after the original data of the machine tool (usually time series data in the time domain, such as vibration, current, etc.) is transformed in the frequency domain (such as Fourier transform), The complex amplitude corresponding to this point. This amplitude includes the amplitude and phase information of the machine tool operation status information at this frequency component (although the square of the amplitude is mainly concerned in the spectrum energy calculation). The vibration and operation characteristics of different machine tool components will reflect frequency domain signals of different amplitudes at different frequencies. By analyzing these amplitude conditions, the frequency characteristics corresponding to potential faults can be found.

[0065] In detail, discrete frequency points are discrete frequency points selected when performing frequency domain analysis on machine tool data. It is its index identifier, which is used to distinguish different frequency points. The value range is usually from arrive These discrete frequency points together constitute the frequency domain range to be analyzed, covering various frequency bands from low frequency to high frequency that may be related to the operating status of the machine tool. Corresponding It reflects the frequency domain characteristics of the machine tool data under this frequency component.

[0066] In detail, as a counting index, it is used to traverse all discrete frequency points, starting from Start, increment each time ,until , which is convenient for accurately referring to information such as the frequency domain signal amplitude corresponding to different frequency points in summation and other operations, so as to complete the calculation of spectrum energy.

[0067] In the embodiment of the present invention, the spectrum bandwidth is generated by an algorithm as follows:

[0068] ;

[0069] in, is the spectrum bandwidth of the target machine tool, is the center frequency of the machine data, is the total number of machine data. is the frequency domain signal amplitude of the machine tool data, is the discrete frequency point of the machine tool data, It is the frequency point index of the machine tool data.

[0070] In detail, An indicator used to describe the width of the frequency distribution range of machine tool data in the frequency domain. This indicator can be used to understand the span of the frequency range where the energy of the machine tool operation signal is concentrated in the frequency domain. A wider spectrum bandwidth may indicate interference from multiple frequency sources during machine tool operation, such as resonance of mechanical parts and harmonic interference of motors. The spectrum bandwidth often behaves differently under different fault modes, so it can help determine whether the machine tool is in a fault state and the approximate type of fault.

[0071] In detail, Refers to the center frequency position of the spectrum energy distribution. It plays a role in locating and measuring the relative frequency offset in the formula for calculating the spectrum bandwidth, and determines which frequency is used as a reference to measure the dispersion of the entire spectrum on the frequency axis. The calculation method can vary depending on the specific definition and application scenario. Common methods include calculating the weighted average frequency and other methods to determine it.

[0072] In detail, The meaning is consistent with that in spectrum energy calculation, that is, the number of discrete frequency points. It is also the number of frequency points determined after discretization processing when performing frequency domain analysis on machine tool data. These frequency points constitute the basis of the entire frequency domain range involved in the subsequent calculation of spectrum bandwidth.

[0073] In detail, The same meaning as in spectrum energy calculation, representing the frequency The frequency domain signal amplitude at the frequency point is the key quantity that reflects the amplitude of the machine tool data at each frequency component in the frequency domain analysis. In the calculation formula of the spectrum bandwidth, it is combined with the frequency point and center frequency The coordinated operation is used to measure the discreteness of the energy distribution of each frequency component relative to the center frequency, and then determine the size of the spectrum bandwidth.

[0074] In detail, Refers to the discrete frequency points selected in frequency domain analysis. Its role and meaning are the same as those in spectrum energy calculation. It is the basic element that constitutes the frequency domain range and carries the frequency domain characteristic information. Corresponding to different frequency domain signal amplitudes , jointly participate in the calculation process of spectrum bandwidth.

[0075] In detail, Also used as a counting index to refer to different discrete frequency points in turn, starting from Start to This ensures that when performing complex operations such as summation, the amplitude and other information related to each frequency point can be accurately corresponded to successfully complete the calculation of the spectrum bandwidth.

[0076] In detail, by calculating the two frequency domain features of spectral energy and spectral bandwidth, and combining them with the previous time domain features such as pulse factor and volatility, it is possible to more comprehensively extract the characteristic information of machine tool data from different angles, providing rich and effective input data for machine tool fault prediction and diagnosis based on deep learning, and improving the accuracy and reliability of fault prediction and diagnosis.

[0077] S2. Generate data features of the target machine tool according to the pulse factor, fluctuation rate, spectrum energy and spectrum bandwidth.

[0078] The pulse factor, fluctuation rate, spectrum energy and spectrum bandwidth are fused to obtain the data characteristics of the target machine tool.

[0079] In detail, the pulse factor, volatility, spectrum energy and spectrum bandwidth extracted from different angles (time domain and frequency domain) are fused to integrate multi-dimensional machine tool operation status information, so that the subsequent deep learning model can comprehensively learn various patterns and laws related to machine tool failures. A single type of feature can only reflect a certain aspect of machine tool operation, while the fused features can provide richer and more complete representations, which helps to improve the accuracy and reliability of fault prediction and diagnosis.

[0080] Specifically, the pulse factor, volatility, spectrum energy, and spectrum bandwidth are directly arranged in a certain order and concatenated into a new feature vector. For example, assuming the pulse factor is , the volatility value is , the value of the spectrum energy is , the spectrum bandwidth is , then the fused feature vector can be expressed as .

[0081] In detail, it can completely retain the information of each original feature without losing the data content of any dimension, making it convenient for subsequent models to directly process and learn the relationship between each feature.

[0082] Similarly, a corresponding weight can be assigned to each feature, and the weight can be determined according to the importance of each feature to machine tool fault prediction and diagnosis. Then, each feature is multiplied by the corresponding weight and then summed to obtain the fused feature value.

[0083] S3. Perform linear transformation on the data features according to the pre-built deep learning model to obtain the primary feature output of the data features.

[0084] The data features are linearly transformed according to the linear transformation algorithm in the pre-built deep learning model to obtain the primary feature output of the data features, where the linear transformation algorithm is:

[0085] ;

[0086] in, is the primary feature output of the data feature, o is the identifier of the data feature, is the hidden layer neuron index in the pre-built deep learning model, is the total number of data features, are the weights from the input layer to the hidden layer in the pre-built deep learning model, is the bias of the hidden layer in the pre-built deep learning model, is the data feature, is the activation function.

[0087] In detail, in this linear transformation process, the main purpose is to map the input raw data features to a new feature space through specific linear combinations and nonlinear transformations (implemented with the help of activation functions), so as to mine deeper information in the data and prepare for further feature processing and final machine tool fault prediction and diagnosis. This process is performed between the input layer and the hidden layer of the pre-built deep learning model.

[0088] In detail, the data features represent the data features of the target machine tool obtained through the steps of feature fusion. These data features reflect the operating status of the machine tool from different angles (such as pulse factor and fluctuation rate in the time domain and spectrum energy and spectrum bandwidth in the frequency domain), and are the basic input information for subsequent linear transformation. For example, suppose the data feature vector obtained after feature fusion is , where each They all carry some key information related to the operation of machine tools.

[0089] In detail, Refers to the number of elements in the input data feature vector, that is, the number of feature dimensions involved in this linear transformation operation. For example, if the fused feature vector has 5 dimensions, then , which determines the number of features that need to participate in the weighted summation in the linear combination operation.

[0090] In detail, It plays a key role in regulating deep learning models, which means that data features to the hidden layer The weight parameter of the connection between neurons. Different weight values ​​determine the influence of each data feature on the corresponding neurons in the hidden layer. The value is obtained through continuous adjustment and optimization during the model training process. For example, a larger The value means that the corresponding feature has a greater impact on the output of the hidden layer neurons. The reasonable setting of weights can enable the model to learn the complex correlation between different features and their importance to the final fault prediction.

[0091] In detail, is the hidden layer The bias term of each neuron plays a role similar to the intercept in a linear function, which enables the activation function of the neuron to perform nonlinear transformation at the appropriate position, increasing the flexibility and expressiveness of the model. The value of is 0. The existence of bias can also give the neuron a basic output value, so as to better fit various complex functional relationships to adapt to the possible nonlinear mapping between machine tool faults and operating characteristics.

[0092] In detail, the activation function is the key to introduce nonlinear factors, because the simple linear combination It can only express linear relationships, but there are often complex nonlinear relationships between machine tool failure modes and operating data characteristics, so an activation function is needed to enhance the nonlinear expression ability of the model.

[0093] In detail, by performing similar operations on the neurons in each hidden layer, the original data features can be converted into new primary feature outputs. These primary feature outputs will serve as inputs for further processing (such as secondary conversion), and gradually explore deep feature representations that are more suitable for machine tool fault prediction and diagnosis.

[0094] S4. Perform secondary transformation on the primary feature output to obtain double transformed features of the primary feature output.

[0095] The primary feature output is secondary transformed using a preset quadratic linear transformation algorithm to obtain a double transformation feature of the primary feature output, wherein the preset quadratic linear transformation algorithm is:

[0096] ;

[0097] in, is the double transformation feature of the primary feature output, is the hidden layer neuron index in the pre-built deep learning model, is the total number of hidden layer neurons in the pre-built deep learning model, is the primary feature output, are the weights from the hidden layer to the output layer in the pre-built deep learning model, is the bias of the output layer.

[0098] In detail, In the process of machine tool fault prediction and diagnosis based on deep learning, the step of secondary transformation of primary feature output is to further explore the more complex and in-depth correlation between the primary features obtained before and the machine tool faults, and generate double transformation features through linear combination and potential nonlinear mapping (depending on whether activation functions and other operations are combined later), so as to make it more conducive to the subsequent accurate judgment of machine tool fault probability, etc., thereby improving the fault prediction and diagnosis capabilities of the entire model.

[0099] In detail, the results obtained through the previous linear transformation steps are the feature values ​​output from the input layer of the deep learning model after the transformation from the input layer to the hidden layer. These primary feature outputs have already undergone the initial nonlinear transformation and feature extraction of the original data features (such as the fused pulse factor, volatility, spectrum energy, and spectrum bandwidth), carrying some abstract information mapped by the machine tool operating status in the hidden layer, and providing basic input data for the secondary transformation.

[0100] In detail, It is the serial number used to identify different neurons in the hidden layer, and its value range is from arrive , this index can clearly refer to the primary features output by each hidden layer neuron, which is convenient for accurately obtaining the corresponding eigenvalues ​​in the weighted sum calculation of the secondary transformation, and then participating in the construction of new dual transformation features.

[0101] In detail, if the hidden layer is set to 10 neurons, then , which means that a different primary feature output will participate in the secondary transformation. The number of different hidden layer neurons will affect the model's ability to extract and express features. Its value usually needs to be determined comprehensively based on factors such as the specific application scenario, data scale, and model complexity.

[0102] In detail, during the model training process, the weight values ​​are continuously adjusted and optimized so that the model can learn the accurate mapping relationship between the hidden layer features and the machine tool faults that need to be predicted in the end, for example, the larger The value means the corresponding For the output layer The output of each neuron has a more significant impact, which reflects the difference in the importance of different primary features in the final fault prediction.

[0103] In detail, It is the final output feature obtained after this secondary transformation, that is, the result after a series of transformations from the input layer to the hidden layer and then to the output layer. It is a further abstraction and integration of the machine tool operation status information, and is more closely related to the machine tool failure. Subsequent operations such as calculating loss functions and comparing with real fault labels can be performed based on these dual transformation features to evaluate the performance of the model and guide the update of model parameters, ultimately achieving accurate prediction of the probability of machine tool failure and corresponding diagnostic functions.

[0104] S5. Update the parameters of the pre-built deep learning model according to the data features and the dual transformation features to obtain an updated deep learning model.

[0105] Calculate the parameter gradient of a preset mean square error loss function with respect to the model parameters of a pre-built deep learning model according to the data features and the dual transformation features;

[0106] The parameters of the pre-built deep learning model are updated according to the parameter gradient to obtain an updated deep learning model.

[0107] In detail, in the machine tool fault prediction and diagnosis method based on deep learning, model parameter updating is a key link. Its purpose is to continuously adjust the model parameters (such as the weights and biases of each layer, etc.) so that the model can better fit the input data characteristics (such as the fused pulse factor, volatility, spectrum energy and spectrum bandwidth, etc.) and the expected output (i.e., the target output corresponding to the actual machine tool fault situation, such as the fault probability or fault category, etc.), thereby improving the accuracy of the model in machine tool fault prediction and diagnosis. Here, the mean square error loss function is used to measure the difference between the model prediction result and the actual situation, and the model parameters are gradually optimized based on the gradient of the loss function with respect to the model parameters and the preset update algorithm, and finally the updated deep learning model is obtained.

[0108] In detail, the mean square error loss function is often used in regression problems to measure the error between the model prediction value and the true value.

[0109] In detail, this function quantifies the overall error of the model by calculating the average of the sum of squares of the difference between the predicted value and the true value of each sample. The smaller the loss function value, the closer the model's prediction result is to the actual situation and the better the model performance.

[0110] Specifically, the parameters of the pre-built deep learning model are updated according to the parameter gradient to obtain an updated deep learning model, including:

[0111] The parameters of the pre-built deep learning model are updated according to the preset update algorithm and parameter gradient to obtain an updated deep learning model, wherein the preset update algorithm is:

[0112] ;

[0113] in, It is The model parameters of the iteration, It is The model parameters of the iteration, is the learning rate, is the number of iterations, It is The parameter gradients of the preset mean squared error loss function with respect to the model parameters of the pre-built deep learning model after iterations.

[0114] In detail, the learning rate is a pre-set positive number. It plays a key regulatory role in the process of updating model parameters and determines the step size of the model parameters along the gradient direction at each iteration. If the learning rate is too large, it may cause the model to "take big steps" when updating parameters, skipping the minimum point of the loss function, making the model unable to converge and even causing the loss function value to become larger and larger, resulting in model divergence; and if the learning rate is too small, although the model parameter update will be in the right direction, the amplitude of each adjustment is too small, which will cause the model to converge too slowly, requiring a large number of iterations to achieve good performance, wasting computing resources and time. The choice of learning rate usually needs to be determined through some experiments and experience. You can first try some common value ranges (such as 0.01, 0.1, 0.001, etc.), and then adjust and optimize according to the performance of the model on the validation set.

[0115] In detail, by iterating the above process of calculating parameter gradients and updating parameters multiple times (the number of iterations will be determined according to a pre-set stopping condition, such as reaching a fixed upper limit of iterations, or when the loss function value of the model on the validation set no longer decreases significantly, etc.), the parameters of the model will be continuously optimized in the direction of reducing the loss function value, and finally an updated deep learning model will be obtained, which can more accurately predict machine tool failures based on the input data features.

[0116] In detail, for example, assuming that the mean square error loss function value between the initial model's prediction results and the actual fault situation is large, with repeated iterative updates, it is observed that the loss function value gradually decreases, indicating that the model is constantly improving. After a certain number of iterations, the model can achieve satisfactory performance on both training data and verification data. At this time, the model is a fully trained and updated deep learning model, which can be used for real-time fault prediction and diagnosis of the target machine tool.

[0117] S6. Collect real-time data of the target machine tool, input the real-time data into the updated deep learning model, obtain the failure probability of the target machine tool, and maintain the target machine tool according to the failure probability.

[0118] In the embodiment of the present invention, collecting real-time data of the target machine tool includes: collecting real-time operation data from the target machine tool, which data includes vibration data, temperature data, noise data, etc. These data are input features that the model relies on to make fault prediction.

[0119] In an embodiment of the present invention, inputting real-time data into an updated deep learning model refers to inputting the collected real-time data into a deep learning model that has been trained and updated.

[0120] In detail, the key to this step is to pass the data into the model and make predictions using the parameters that have been adjusted inside the model.

[0121] Furthermore, the updated deep learning model outputs the probability of machine tool failure based on the input real-time data.

[0122] In an embodiment of the present invention, maintaining a target machine tool according to the failure probability includes: determining whether to maintain the machine tool according to the obtained failure probability.

[0123] For example, when the probability of failure is high, a maintenance or shutdown inspection process may be triggered, thereby preventing machine tool failure from occurring, reducing downtime and repair costs.

[0124] It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0125] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method and technology of using digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0126] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A machine tool fault prediction and diagnosis method based on deep learning, characterized in that: The following steps are involved: (1) Collect machine tool data to obtain pulse factor, fluctuation rate, spectrum energy and spectrum bandwidth; (2) Generate data features of the target machine tool based on pulse factor, fluctuation rate, spectrum energy, and spectrum bandwidth; (3) Perform linear transformation on data features according to the pre-built deep learning model to obtain primary feature output; (4) Perform secondary transformation on the primary feature output to obtain double transformed features of the primary feature output; (5) Update the parameters of the pre-built deep learning model according to the data features and the dual transformation features to obtain an updated deep learning model; (6) Collect the real-time data of the target machine tool and input the real-time data into the updated deep learning model to obtain the failure probability of the target machine tool.

2. The machine tool fault prediction and diagnosis method based on deep learning according to claim 1 is characterized in that: In step (1), the calculation formula of the pulse factor is as follows: ； in, is the pulse factor of the target machine tool, It is the machine data. is the maximum number in the pre-collected machine data sequence, is the standard deviation of the pre-collected machine data series.

3. The machine tool fault prediction and diagnosis method based on deep learning according to claim 1, characterized in that: In step (1), the volatility is calculated as follows: ； in, is the volatility of the target machine tool, It is the machine data. is the maximum number in the pre-collected machine data sequence, is the minimum number in the pre-collected machine data sequence.

4. The machine tool fault prediction and diagnosis method based on deep learning according to claim 1, characterized in that: In step (1), the calculation formula of spectrum energy is as follows: ； in, is the spectral energy of the target machine tool, is the total number of machine data. is the frequency domain signal amplitude of the machine tool data, is the discrete frequency point of the machine tool data, It is the frequency point index of the machine tool data.

5. The machine tool fault prediction and diagnosis method based on deep learning according to claim 1, characterized in that: In step (1), the spectrum bandwidth is calculated as follows: ； in, is the spectrum bandwidth of the target machine tool, is the center frequency of the machine data, is the total number of machine data. is the frequency domain signal amplitude of the machine tool data, is the discrete frequency point of the machine tool data, It is the frequency point index of the machine tool data.

6. The machine tool fault prediction and diagnosis method based on deep learning according to claim 1, characterized in that: In step (2), after the pulse factor, fluctuation rate, spectrum energy and spectrum bandwidth are fused, the data features of the target machine tool are obtained.

7. The machine tool fault prediction and diagnosis method based on deep learning according to claim 1, characterized in that: In step (3), the data features are linearly transformed according to the linear transformation algorithm in the pre-built deep learning model to obtain the primary feature output of the data features, wherein the linear transformation algorithm is: ； in, is the primary feature output of the data feature, o is the identifier of the data feature, is the hidden layer neuron index in the pre-built deep learning model, is the total number of data features, are the weights from the input layer to the hidden layer in the pre-built deep learning model, is the bias of the hidden layer in the pre-built deep learning model, is the data feature, is the activation function.

8. The machine tool fault prediction and diagnosis method based on deep learning according to claim 1, characterized in that: In step (4), the primary feature output is secondary transformed using a preset quadratic linear transformation algorithm to obtain a double transformation feature of the primary feature output, wherein the preset quadratic linear transformation algorithm is: ； in, is the double transformation feature of the primary feature output, is the hidden layer neuron index in the pre-built deep learning model, is the total number of hidden layer neurons in the pre-built deep learning model, is the primary feature output, are the weights from the hidden layer to the output layer in the pre-built deep learning model, is the bias of the output layer.

9. The machine tool fault prediction and diagnosis method based on deep learning according to claim 1, characterized in that: The specific process of step (5) is as follows: (5-1) Calculating the parameter gradient of the preset mean square error loss function with respect to the model parameters of the deep learning model according to the data features and the dual transformation features; (5-2) The parameters of the pre-built deep learning model are updated according to the parameter gradient and the preset update algorithm to obtain an updated deep learning model.

10. The machine tool fault prediction and diagnosis method based on deep learning according to claim 9, characterized in that: In step (5-2), the preset update algorithm is: ； in, It is The model parameters of the iteration, It is The model parameters of the iteration, is the learning rate, is the number of iterations, It is The parameter gradient of the preset mean square error loss function with respect to the model parameters of the deep learning model after iterations.

Images