Machine abnormity detection method and abnormal machine detection equipment

Abstract

The invention discloses a machine abnormity detection method and abnormal machine detection equipment. The anomaly detection method comprises the steps of obtaining working parameters of a machine; the working parameters are input into a preset trained detection model to be processed, a processing result is output, and output of the trained detection model is at least determined by a residual function, an identical mapping, a first dynamic weight related to the residual function and a second dynamic weight related to the identical mapping; and judging whether the machine is an abnormal machine according to the processing result.

Description

Technical Field

[0001] This application relates to the field of machine tool testing technology, specifically to a method and equipment for detecting abnormalities in machine tools. Background Technology

[0002] During machining using Computer Numerical Control (CNC) machine tools, malfunctions can occur due to issues such as tool failure, transmission mechanism failure, program failure, or electrical and control system failures, potentially leading to a decrease in the yield rate of machined products. Typically, multiple sensors are required on the CNC machine tool to detect abnormal conditions. Different sensor types or detection methods are needed for different machine models and levels of aging, resulting in high costs and low accuracy for CNC machine tool malfunction detection. Summary of the Invention

[0003] To address the issues of high cost and low accuracy in CNC machine tool anomaly detection, this application provides an anomaly detection method and anomaly detection equipment for CNC machine tools.

[0004] This application provides a method for detecting anomalies in a machine, comprising: acquiring the machine's operating parameters; inputting the operating parameters into a trained detection model for processing and outputting the processing result, wherein the output processing result of the trained detection model is determined at least by a residual function, an identity mapping, a first dynamic weight associated with the residual function, and a second dynamic weight associated with the identity mapping; and determining whether the machine is an abnormal machine based on the processing result.

[0005] In some embodiments, the anomaly detection method further includes: preprocessing the operating parameters; The working parameters are input into the trained detection model for processing to output the processing results, including: The preprocessed working parameters are input into the trained detection model for further processing to output the processing results.

[0006] In some embodiments, the preprocessing includes at least one of noise reduction, normalization, and data normalization.

[0007] In some implementations, the trained detection model includes a first processing layer, a second processing layer, and a position vector. The first processing layer includes a CNN layer, an upsampling layer, and a dynamic weight acquisition layer. The position vector is configured to sort other inputs to the second processing layer besides the position vector. The step of inputting the working parameters into the trained detection model for processing to output a processing result, or the step of inputting the preprocessed working parameters into the trained detection model for processing to output a processing result, includes: The dynamic weight acquisition layer processes the working parameters to obtain the first dynamic weight and the second dynamic weight; the CNN layer processes the working parameters to obtain a first intermediate parameter, which is determined by the residual function in the CNN layer; the upsampling layer performs a constant mapping on the working parameters to obtain a second intermediate parameter; a third intermediate parameter is obtained based on the first intermediate parameter, the first dynamic weight, the second intermediate parameter, and the second dynamic weight; the second processing layer processes the third intermediate parameter and the position vector to obtain the processing result.

[0008] In some embodiments, the second processing layer includes an encoding layer, a decoding layer, and a fully connected layer; the second processing layer processes the third intermediate parameter and the position vector to obtain the processing result, including: The encoding layer extracts features from the third intermediate parameter after sorting the position vectors; the decoding layer restores the feature extraction results from the encoding layer to obtain the restored result; the fully connected layer reshapes the restored result so that the processing result output by the second processing layer has the same data form as the working parameters.

[0009] In some embodiments, the operating parameters are obtained from at least one processing process in the history of the machine tool, and the operating parameters include at least one of the following: current value, voltage value, temperature, torque, speed, power, load and feed rate of the machine tool in the operating state.

[0010] In some implementations, determining whether the machine is abnormal based on the processing result includes: simultaneously inputting the processing result and the working parameters into a loss function to output an evaluation value; if the evaluation value is within a preset evaluation value threshold, determining the machine is abnormal; if the evaluation value is outside the preset evaluation value threshold, determining the machine is healthy.

[0011] In some embodiments, before processing the operational parameters by inputting them into the trained detection model, the anomaly detection method further includes: acquiring historical anomaly parameters of multiple abnormal machines to form a training operational parameter set, wherein the historical anomaly parameters are acquired through at least one processing operation of the abnormal machine in an abnormal state, and the historical anomaly parameters include at least one of the current value, voltage value, temperature, torque, speed, power, load, and feed rate of the abnormal machine in the abnormal operating state; and inputting the training operational parameter set into an initial training model for training to obtain the trained detection model.

[0012] In some implementations, the initial training model includes a first module and a second module. The first module includes a CNN module, an upsampling module, and a dynamic weight acquisition module. The second module includes an encoding module, a decoding module, and a fully connected module. Position vectors are configured to sort other inputs to the second module besides the position vectors. The step of inputting the training working parameter set into the initial training model for training to obtain the trained detection model includes: inputting at least a portion of the historical anomaly parameters from the training working parameter set into the CNN module, the upsampling module, and the dynamic weight acquisition module respectively to obtain first intermediate training parameters; the encoding module extracts features from the sorted position vectors of the first intermediate training parameters; and the decoding module extracts features from the encoded... The feature extraction results of the module are restored; the fully connected module shapes the restored results of the decoding module so that the training results output by the second module have the same data form as the historical anomaly parameters; the training results and the historical anomaly parameters are input into a loss function to output a feedback value; if the feedback value is greater than the feedback value threshold, the initial training model is updated according to the feedback value; and at least a portion of the remaining data of the training working parameter set is input into the updated initial training model, and the updated initial training model is used to perform the steps from obtaining the first intermediate training parameters to updating the initial training model according to the feedback value, until the feedback value is less than or equal to the feedback value threshold, then the current initial training model is used as the detection model after training is completed.

[0013] Secondly, this application provides an abnormal machine tool detection device, which includes a memory and a processor. The memory is used to store instructions. The instructions stored in the memory are executed by the processor to implement the abnormal detection method described in any of the above embodiments.

[0014] The present application discloses an anomaly detection method and anomaly detection device for machines. It utilizes a trained detection model to process the operating parameters of the machine to be detected, obtaining a processing result related to whether the machine is in an abnormal state. This processing result is determined at least by the residual function, the identity mapping, a first dynamic weight associated with the residual function, and a second dynamic weight associated with the identity mapping. In this case, only the machine's operating parameters are needed for anomaly detection, eliminating the need for additional data specifically collected for anomaly detection. This method can be used to detect and analyze machines of different models and aging levels, and the data used in the analysis is comprehensive, resulting in low detection cost and high accuracy. Through the synergistic effect of dynamic weights, the residual function, and the identity mapping, the network degradation and gradient vanishing problems existing in neural networks are effectively solved. Furthermore, the participation of dynamic weights gives the trained detection model stronger generalization ability. The trained detection model can set the first dynamic weight and the second dynamic weight according to the specific data content of the working parameters. It can effectively capture features through the residual function and reduce costs and alleviate information redundancy through identity mapping. The trained detection model has strong robustness and is applicable to different models of machines.

[0015] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0016] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, wherein: Figure 1 This is a flowchart illustrating the abnormality detection method for a machine according to some embodiments of this application; Figure 2 This is a flowchart illustrating the abnormality detection method for a machine according to some embodiments of this application; Figure 3 This is a flowchart illustrating the abnormality detection method for a machine according to some embodiments of this application; Figure 4 This is a flowchart illustrating the abnormality detection method for a machine according to some embodiments of this application; Figure 5 This is a flowchart illustrating the abnormality detection method for a machine according to some embodiments of this application; Figure 6 This is a flowchart illustrating the abnormality detection method for a machine according to some embodiments of this application; Figure 7 This is a flowchart illustrating the abnormality detection method for a machine according to some embodiments of this application; Figure 8This is a flowchart illustrating the abnormality detection method for a machine according to some embodiments of this application; Figure 9 This is a structural schematic diagram of an abnormal machine detection device according to some embodiments of this application.

[0017] The reference numerals in the detailed embodiments are as follows: 100 abnormal machine detection devices; Memory 10; Processor 30. Detailed Implementation

[0018] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the specific embodiments of this application are described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0019] In the description of this application, it should be understood that the terms "center", "length", "upper", "lower", "front", "rear", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0020] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0021] In this application, unless otherwise expressly specified and limited, the terms "installation," "connection," "joining," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise expressly limited. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0022] In this application, unless otherwise expressly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "on top of," and "over" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0023] During machining using Computer Numerical Control (CNC) machine tools, malfunctions can occur due to issues such as tool malfunctions, transmission mechanism malfunctions, program malfunctions, or electrical and control system malfunctions, potentially leading to a decrease in the yield rate of machined products. Typically, multiple sensors are required on the CNC machine tool to detect abnormal conditions. Different sensor models or detection methods are needed for different machine tool models and aging levels, resulting in high costs and low accuracy for CNC machine tool anomaly detection. To address these problems, this application provides a machine tool anomaly detection method (…). Figures 1 to 8 (as shown) and abnormal machine detection equipment 100 ( Figure 9 (As shown).

[0024] Please refer to Figure 1 This application provides a method for detecting machine anomalies, the method comprising: 01: Obtain the machine's operating parameters; 07: Input the working parameters into the pre-trained detection model for processing and output the processing result. The output of the pre-trained detection model is determined at least by the residual function, the identity mapping, the first dynamic weight related to the residual function, and the second dynamic weight related to the identity mapping; and 09: Determine whether the machine is abnormal based on the processing results; The process of determining whether a machine is abnormal based on the processing results includes: assigning a residual function to the first dynamic weights, assigning an identity mapping to the second dynamic weights to optimize the processing results output by the trained detection model, inputting the optimized processing results and working parameters into the loss function to output an evaluation value, and comparing the evaluation value with a preset evaluation value threshold to determine whether the machine is abnormal.

[0025] Correspondingly, please combine Figure 9This application provides an abnormal machine detection device 100. The abnormal machine detection device 100 includes a memory 10 and a processor 30. The memory 10 is used to store instructions. The instructions stored in the memory 10 are executed by the processor 30 to implement the abnormal detection methods in 01, 07, and 09. More specifically, the processor 30 is used to: acquire the machine's operating parameters; input the operating parameters into a preset, trained detection model for processing, and output the processing result, wherein the output of the trained detection model is determined at least by a residual function, an identity mapping, a first dynamic weight related to the residual function, and a second dynamic weight related to the identity mapping; and determine whether the machine is an abnormal machine based on the processing result.

[0026] Specifically, a Computer Numerical Control (CNC) machine tool is a machining tool that can be controlled by pre-programmed computer software. Users can control the working parameters of the cutting tools and the machining path through pre-edited programs, thereby performing precise cutting, carving, drilling, and other machining processes on substrates (such as metals, plastics, and wood). CNC machine tools (hereinafter referred to as "machine tools") have many advantages such as high precision, good repeatability, high machining efficiency, high automation, and high flexibility, and are widely used in various machining processes. However, during long-term use, machine tools may malfunction, such as tool malfunctions, transmission mechanism malfunctions, program malfunctions, or electrical and control system malfunctions, which reduces the yield rate of machining. This application provides a method for detecting machine tool malfunctions. This method only requires the machine tool's working parameters to detect malfunctions, without the need for additional sensors to collect data specifically for malfunction detection, and has many advantages such as low detection cost and high detection accuracy. Simultaneously, the first dynamic weight is assigned to the residual function, and the second dynamic weight is assigned to the identity mapping to optimize the processing result output by the trained detection model. The optimized processing result and the working parameters are simultaneously input into the loss function to output an evaluation value. The evaluation value is compared with a preset evaluation value threshold to determine whether the machine is abnormal. The trained detection model can set the first dynamic weight and the second dynamic weight according to the specific data content of the working parameters. It can effectively capture features through the residual function and reduce costs and alleviate information redundancy through the identity mapping. It can be applied to different models of machines and significantly reduces the probability of abnormal machines being misjudged, thus improving the accuracy of detection.

[0027] The abnormal machine detection device 100 is a device used to detect whether a machine has any abnormalities. The memory 10 stores instructions corresponding to the abnormality detection method, and the processor 30 executes the instructions stored in the memory 10 to enable the abnormal machine detection device 100 to implement the abnormality detection method. The memory 10 and the processor 30 are electrically connected and connected to various parts of the entire abnormal machine detection device 100 via various interfaces and lines. The processor 30 executes methods 01, 07, and 09 by running or loading instructions stored in the memory 10 and by calling data stored in the memory 10, thereby realizing the abnormality detection of the machine.

[0028] During the processing of the substrate, the machine tool executes preset processing steps with certain operating parameters. These operating parameters include, but are not limited to, physical quantities, state values, and logical conditions during the processing. For example, operating parameters may include the machine tool's current value, voltage value, temperature, torque, speed, power, and feed rate. The machine tool's operating parameters reflect its state (e.g., operating status, health status). Therefore, the anomaly detection method of this application uses the acquired machine tool operating parameters for anomaly detection, eliminating the need for a dedicated detector to detect anomaly-specific data. This reduces the complexity of the machine tool system, lowers the cost of anomaly detection, and provides a comprehensive range of operating parameters for analysis, allowing these parameters to work synergistically to better reflect the machine tool's state.

[0029] Furthermore, the processor 30 can perform anomaly detection based on the acquired working parameters from one or more processing operations. Specifically, the processor 30 inputs the working parameters into a pre-trained detection model. The trained detection model is a model that analyzes the health status of the machine based on the working parameters. The trained detection model can be a fixed model set at the factory or a variable model that can be adjusted and optimized by the operator.

[0030] The trained detection model employs at least two processing methods—residual function and identity mapping—in its handling of operating parameters. The residual function is used for cross-modal fusion of multi-source, heterogeneous features, enabling the trained detection model to analyze and process operating parameters of different machine types and states. The identity mapping preserves features from the original input information, preventing information loss. Simultaneously, the trained detection model assigns a first dynamic weight to the residual function and a second dynamic weight to the identity mapping. The first dynamic weight controls and determines the weight of the residual function in the trained detection model, while the second dynamic weight controls and determines the weight of the identity mapping, allowing the trained detection model to be flexibly adjusted based on specific operating parameter data.

[0031] It can be understood that the residual function, the identity mapping, the first dynamic weight related to the residual function, and the second dynamic weight related to the identity mapping determine the specific processing method of the trained detection model on the input data. Therefore, the processing result output by the trained detection model after processing the working parameters is determined by the above. Furthermore, the processor 30 can determine whether the machine is an abnormal machine based on the processing result output by the trained detection model.

[0032] The machine tool anomaly detection method and anomaly machine tool detection device 100 of this application utilize a trained detection model to process the operating parameters of the machine tool to be detected, so as to obtain the processing result related to whether the machine tool is in an abnormal state. The processing result is determined at least by the operating parameter processing result obtained by the residual function, the original operating parameters obtained by the identity mapping, the first dynamic weight related to the residual function, and the second dynamic weight related to the identity mapping. At this time, only the operating parameters of the machine tool are needed to perform anomaly detection, without the need to collect additional data specifically for anomaly detection. Machine tools of different models and different aging levels can be detected and analyzed using this method, and the data used in the analysis process is relatively comprehensive, with low detection cost and high detection accuracy.

[0033] Please refer to Figure 1 and Figure 2 In some implementations, the anomaly detection method further includes: 03: Preprocess the working parameters; 07: Input the working parameters into the preset trained detection model for processing, and output the processing results, including: 071: Input the preprocessed working parameters into the preset trained detection model for processing, and output the processing results.

[0034] Furthermore, please combine Figure 9 The processor 30 is also used to execute the methods in 03 and 071. Specifically, the processor 30 is configured to: preprocess the working parameters; input the preprocessed working parameters into a preset trained detection model for processing, and output the processing results.

[0035] Specifically, it should be noted that due to differences in model, service life, aging degree, working parameters, acquisition frequency, and processing cycle time (CT), the working parameters acquired by the machine may contain abnormal extreme values, noise, and inconsistent data lengths. Therefore, directly using the raw working parameters for analysis may affect the stability and reliability of the output results of the trained detection model. The processor 30 needs to preprocess the raw working parameters acquired before inputting the processed structure into the trained detection model for analysis.

[0036] At this point, method 07, which inputs the working parameters into a preset, trained detection model for processing and outputs the processing result, is further optimized into method 071, whereby processor 30 first preprocesses the working parameters and then inputs the preprocessed working parameters into a preset, trained detection model for processing and outputs the processing result.

[0037] Therefore, in the anomaly detection method of this application, the processor 30 preprocesses the collected raw working parameters to eliminate the impact of noise and other problems on data quality. At this time, the quality of the data input to the trained detection model is better, and the processing results output by the trained detection model can more realistically and effectively reflect the status of the machine. The accuracy and reliability of the anomaly detection method are better.

[0038] Please refer to Figure 1 and Figure 2 In some implementations, preprocessing includes at least one of denoising, normalization, and data normalization.

[0039] Specifically, in the above embodiments, the processing of the original working parameters includes, but is not limited to, one or more combinations of denoising, normalization, and data warping. Denoising is used to remove noise from the original working parameters, and may include, but is not limited to, one or more combinations of Fourier denoising, moving average filtering, median filtering, and Savitzky-Golay filtering. Normalization is used to remove extreme values. Data warping is used to process data with inconsistent lengths to enable collaborative analysis. For example, the processor 30 performs data warping through Dynamic Time Warping (DTW), specifically by using dynamic programming techniques to flexibly align two data sequences with inconsistent lengths.

[0040] Therefore, in the anomaly detection method of this application, the processor 30 preprocesses the collected raw working parameters through at least one of denoising processing, normalization processing and data regularization processing to eliminate the influence of noise and other problems on data quality. At this time, the data quality of the detection model after training is better, and the processing results output by the detection model after training can more realistically and effectively reflect the status of the machine. The accuracy and reliability of the anomaly detection method are better.

[0041] Please refer to Figures 1 to 3In some implementations, the trained detection model includes a first processing layer and a second processing layer. The first processing layer includes a CNN layer, an upsampling layer, and a dynamic weight acquisition layer. The position vector is configured to sort the inputs to the second processing layer other than the position vector. 07: Input the working parameters into the preset trained detection model for processing to output the processing result, or, 071: Input the preprocessed working parameters into the preset trained detection model for processing to output the processing result, including: 0711: The dynamic weight acquisition layer processes the working parameters to obtain the first dynamic weight and the second dynamic weight; 0713: The CNN layer processes the working parameters to obtain the first intermediate parameter, which is determined by the residual function in the CNN layer; 0715: The upsampling layer performs a constant mapping on the working parameters to obtain the second intermediate parameter; 0717: Obtain the third intermediate parameter based on the first intermediate parameter, the first dynamic weight, the second intermediate parameter, and the second dynamic weight; 0719: The second processing layer processes the third intermediate parameter and the position vector to obtain the processing result.

[0042] Furthermore, please combine Figure 9 The processor 30 is also used to execute the methods in 0711, 0713, 0715, 0717, and 0719. Specifically, the processor 30 is configured as follows: a dynamic weight acquisition layer processes the working parameters to obtain a first dynamic weight and a second dynamic weight; a CNN layer processes the working parameters to obtain a first intermediate parameter, which is determined by the residual function in the CNN layer; an upsampling layer performs a constant mapping on the working parameters to obtain a second intermediate parameter; a third intermediate parameter is obtained based on the first intermediate parameter, the first dynamic weight, the second intermediate parameter, and the second dynamic weight; and a second processing layer processes the third intermediate parameter and the position vector to obtain a processing result.

[0043] Specifically, the first processing layer and the second processing layer are the internal structures of the trained detection model, which are distributed sequentially and work together to achieve the detection process of the working parameters. After the working parameters are input into the trained detection model, they are first input into the first processing layer. The intermediate parameters obtained by the first processing layer are then further input into the second processing layer for analysis and processing, and the second processing layer outputs the processing results.

[0044] More specifically, in the first processing layer, the CNN layer is the core feature extractor in the trained detection model. The CNN layer includes a residual function and can analyze and perceive the input working parameters (or preprocessed working parameters, hereinafter referred to as working parameters) through a series of learnable convolutional kernels. The upsampling layer is the part of the first processing layer used to implement the identity mapping. The dynamic weight acquisition layer is the part of the first processing layer used to determine the weights corresponding to the distributions of the CNN layer and the upsampling layer (i.e., the weights corresponding to the residual function and the identity mapping).

[0045] The first dynamic weight and the second dynamic weight are weight values ​​dynamically generated by the dynamic weight acquisition layer based on the input data. The first dynamic weight controls the contribution weight of the residual function in the CNN layer, and the second dynamic weight controls the contribution weight of the constant mapping in the upsampling layer. In this case, the first processing layer has a strategy that can adaptively adjust feature fusion according to the characteristics of the input. In some implementations, the sum of the first dynamic weight and the second dynamic weight is 1 to ensure that the third intermediate parameter of the output is a normalized weighted combination. It can be understood that the closer the first dynamic weight is to 1, the more the first processing layer emphasizes feature transformation, and it is suitable for applications processing complex data; the closer the second dynamic weight is to 1, the more the first processing layer emphasizes preserving the original input, and it is suitable for applications that directly transmit information.

[0046] More specifically, the first dynamic weight The calculation formula is:

[0047] in, It is a time-series working parameter sequence (i.e., a sequence of parameters corresponding to a certain feature in the input working parameters according to timestamps). The number of features in the working parameters. The parameters for the dynamic weight acquisition layer are obtained from the dynamic residuals based on the input. This represents the intercept, which is mainly used to adjust and optimize the performance of the algorithm.

[0048] More specifically, the second dynamic weight The calculation formula is:

[0049] The calculation process for the second dynamic weight is the same as above, and will not be repeated here.

[0050] For example, consider machine A (an older machine) and machine B, both of which are in good working order. Taking the first processing layer as a simplified model, and using only the spindle temperature as a single operating parameter, the spindle temperature in the timing parameter sequence of machine A is... The spindle temperature in the timing parameter sequence of machine tool B is [30,35,30,50,55]. The values ​​are: [20, 25, 20, 40, 45], which are learned by the dynamic weight acquisition layer based on the working parameters. For [0,0,0,0,0], the values ​​of machine A and machine B are... They are [0.01, 0.01, ... 0.01, [0.01, 0.01] and [-0.01, 0.01, 0, 0, 0]. At this time, the second dynamic weight of machine A is... = Based on the fact that the sum of the first dynamic weight and the second dynamic weight is 1, we know that the first dynamic weight is 1 - 0.4 = 0.6; the second dynamic weight of machine B is... = The first dynamic weight is 1-0.05=0.95.

[0051] The CNN layer performs feature extraction and multi-layer CNN convolution on the working parameters to extract transient features of the working parameters and outputs the analysis results as the first intermediate parameter. This first intermediate parameter contains data reflecting the machine's state and is determined by the residual function in the CNN layer. More specifically, the residual function is used to perform a non-linear transformation on the input working parameters (such as calculating the output in a convolutional or fully connected layer) to learn the residual (i.e., the difference) between the input and the target output. The residual function can be used to extract higher-order features, abstract features, and solve complex mapping problems.

[0052] The upsampling layer performs a constant mapping on the working parameters and outputs them as the second intermediate parameters. In essence, the upsampling layer preserves the original data content of the working parameters. Therefore, the upsampling layer can retain shallow features in the working parameters, preventing information loss and ensuring that the gradient in the first processing layer can propagate back through a shortcut path, mitigating the gradient vanishing problem that may occur in the first processing layer. For example, the gradient vanishing problem refers to the potential loss of information if, among many working parameters, the spindle temperature exhibits an abnormal instantaneous fluctuation.

[0053] Furthermore, the first processing layer integrates the first intermediate parameter, the first dynamic weight, the second intermediate parameter, and the second dynamic weight to obtain the third intermediate parameter. Therefore, through the synergistic effect of dynamic weights, residual functions, and identity mapping, the first processing layer effectively solves the network degradation and gradient vanishing problems existing in neural networks. The participation of dynamic weights also gives the trained detection model stronger generalization ability. The trained detection model can set the first and second dynamic weights according to the specific data content of the working parameters. It effectively captures features through the residual function and reduces costs and alleviates information redundancy through the identity mapping. The trained detection model has strong robustness and is applicable to different types of machines. It can be understood that the third intermediate parameter is the output of the first processing layer when the input is the working parameters.

[0054] Continuing the example above, because machine A is an older model, its spindle temperature is higher under the same process and machining technology. If the acquisition process of the third intermediate parameter does not consider the first and second dynamic weights, the abnormal spindle temperature of machine A may cause the first intermediate parameter output by the CNN layer to be abnormal, potentially leading to machine A being detected as an abnormal machine. However, by introducing dynamic weights through the dynamic weight acquisition layer, the processor 30 dynamically reduces the first dynamic weight based on the higher spindle temperature in the input working parameters of machine A, thus making the generated third intermediate parameter more accurately reflect the actual situation of machine A and reducing the impact of the older machine causing the higher spindle temperature on the anomaly detection method.

[0055] More specifically, the residual function in the CNN layer is F(X) = The mapping relationship of constant mapping in the upsampling layer corresponds to Identity(x) = At this point, if the first and second dynamic weights are not considered and a direct residual connection is performed directly, then the third intermediate parameter of machine A is F( )+Identity( The third intermediate parameter of machine A is 165, with a value of 220. However, after introducing dynamic weights by setting a dynamic weight acquisition layer, the formula for calculating the third intermediate parameter of machine A is... F(X)+ Identity(x), that is, the third intermediate parameter of machine A at this time is 0.4F ( )+0.6Identity( The third intermediate parameter of machine B is 143.25, with a value of 128. The difference between machine A (the normal old machine) and machine B (the normal new machine) decreases from (165-220) / 220=25% to (143.25-128) / 143.25=10.65%, significantly reducing the probability of machine A being misclassified as an abnormal machine. The accuracy of the analysis in the first processing layer is significantly improved.

[0056] The position vector is used to sort all inputs other than the position vector to the second processing layer. When the third intermediate parameter and the position vector are used as inputs, the position vector sorts the third intermediate parameter so that the data in the third intermediate parameter are arranged in a predetermined order, allowing the second processing layer to further analyze and process the sorted third intermediate parameter. The processing result of the second processing layer on the third intermediate parameter is the output of the trained detection model, which is used to analyze the status of the machine.

[0057] Therefore, in the anomaly detection method of this application, the trained detection model processes the input working parameters through a CNN layer, an upsampling layer, and a dynamic weight acquisition layer, and outputs a third intermediate parameter. The third intermediate parameter, together with the position vector, is then input to the second processing layer to obtain the processing result. At this point, the processing result can better reflect the information related to the machine's state contained within the working parameter. The trained detection model has a good ability to interpret data and has good accuracy and reliability in detecting machine anomalies.

[0058] Please refer to Figures 1 to 4 In some implementations, the second processing layer includes an encoding layer, a decoding layer, and a fully connected layer. 0719: The second processing layer processes the third intermediate parameter and the position vector to obtain a processing result, including: 07191: The encoding layer extracts features from the third intermediate parameter after sorting the position vectors; 07193: The decoding layer restores the feature extraction results from the encoding layer to obtain the restored result; 07195: The fully connected layer shapes the restoration result so that the processing result output by the second processing layer has the same data format as the working parameters.

[0059] Furthermore, please combine Figure 9 The processor 30 is also used to execute the methods in 07191, 07193, and 07195. Specifically, the processor 30 is configured to: extract features from the third intermediate parameters after the position vectors are sorted by the encoding layer; restore the feature extraction results of the encoding layer to obtain the restored results; and shape the restored results by the fully connected layer so that the processing results output by the second processing layer have the same data form as the working parameters.

[0060] Specifically, in the above implementation, the encoding layer is the part of the second processing layer used for feature extraction from the input. The core task of the encoding layer is to understand and represent the input sequence, and the encoding layer can extract the input into a set of feature vectors. At this time, the encoding layer can analyze the third intermediate parameter, which is sorted by the position vector and has a specific order of internal data, thereby extracting features and generating feature extraction results (feature vectors), and then sending the feature extraction results to the decoding layer.

[0061] The decoding layer is the part of the second processing layer used to restore the data from the feature extraction results generated by the encoding layer. The core task of the decoding layer is to work with the encoding layer to generate the restored result based on the feature vectors extracted by the encoding layer, ensuring the rationality of the generation process and its fidelity to the input feature vectors. At this point, the decoding layer can decode and restore the feature extraction results generated by the encoding layer, thus generating the restored result, and then feed the restored result to the fully connected layer.

[0062] The fully connected layer is the part of the second processing layer used to shape the restored result generated by the decoding layer. The core task of the fully connected layer is to shape the input content according to a preset data format, such as adjusting data length or sorting data. At this point, the fully connected layer shapes the data format in the restored result generated by the decoding layer and outputs the shaped result as the processing result. It can be understood that the output of the fully connected layer is the output of the second processing layer. The processed result output by the fully connected layer, after shaping, has the same data format as the initial input working parameters of the first processing layer.

[0063] Therefore, in the anomaly detection method of this application, the second processing layer processes the third intermediate parameter after sorting the input position vector through an encoding layer, a decoding layer, and a fully connected layer, and outputs the processing result. The processing result has the same data format as the input working parameter. Since the second processing layer can effectively extract the features related to the machine state contained in the third intermediate parameter, the processing result can better reflect the information related to the machine state contained in the working parameter. The trained detection model has a good ability to interpret data and has good accuracy and reliability in detecting machine anomalies.

[0064] Please refer to Figures 1 to 4 In some implementations, the operating parameters are obtained from at least one historical machining process of the machine tool using existing sensors. The operating parameters include at least one of the following: current, voltage, temperature, torque, speed, power, load, and feed rate of the machine tool during operation.

[0065] Specifically, in the above embodiment, the working parameters are specific parameter values ​​from at least one processing step in the historical processing of the machine tool under test, and these parameter values ​​are collected from existing inspection pieces required for the machine tool to perform the basic processing. In this case, the machine tool does not need to be equipped with additional inspection pieces that are dedicated to the anomaly detection method and do not participate in the basic processing of the machine tool.

[0066] When the machine tool is in operation, at least one of the following parameters—current, voltage, temperature, torque, speed, power, load, and feed rate—can be used as operating parameters for anomaly detection, i.e., input into the trained detection model. At this point, the trained detection model can utilize a relatively rich variety of data, and the aforementioned operating parameters can reflect the machine tool's status from multiple perspectives.

[0067] Therefore, in the machine tool anomaly detection method of this application, the processor 30 only needs to use the data collected by the existing detection parts of the machine tool in at least one processing process in the historical processing process as working parameters, without the need to collect additional data specifically for anomaly detection. The machine tool has low structural complexity, low production cost and maintenance cost, and the data used in the analysis process is more comprehensive, resulting in high detection accuracy of the anomaly detection method.

[0068] Please refer to Figure 1 and Figure 5 In some implementations, 09: Determining whether a machine is abnormal based on the processing result includes: 091: Input the processing results and working parameters into the loss function to output an evaluation value. If the evaluation value is within the preset evaluation value threshold, the machine is determined to be an abnormal machine; if the evaluation value is outside the preset evaluation value threshold, the machine is determined to be a healthy machine.

[0069] Furthermore, please combine Figure 9 The processor 30 is also used to execute the method in 091. Specifically, the processor 30 is configured to: simultaneously input the processing result and the working parameters into the loss function to output an evaluation value; if the evaluation value is within a preset evaluation value threshold, determine that the machine is an abnormal machine; if the evaluation value is outside the preset evaluation value threshold, determine that the machine is a healthy machine.

[0070] Specifically, in the above embodiments, when the processor 30 obtains a processing result with the same data form as the working parameters, the processor 30 is able to perform collaborative analysis and comparison of the processing result and the working parameters.

[0071] It's important to note that because the trained detection model is trained on data with specific properties (such as normal or abnormal data), when the operating parameters differ little from the training data, the trained model can effectively reconstruct the data. In this case, the output (processed result) of the trained detection model is close to the input (operating parameters), resulting in a small loss value (i.e., the evaluation value). However, when the operating parameters differ significantly from the training data, the model cannot effectively reconstruct the data, leading to a large discrepancy between the output (processed result) and the input (operating parameters), and a significantly higher evaluation value.

[0072] At this point, by setting an evaluation value threshold, the evaluation value can directly reflect the degree to which the data deviates from the training data type, thus serving as an anomaly indicator to assess the status of the machine. If the evaluation value is within the preset evaluation value threshold, the processor 30 determines that the model reconstruction process is abnormal and identifies the machine as an abnormal machine; if the evaluation value is outside the preset evaluation value threshold, the processor 30 determines that the model reconstruction process is normal and identifies the machine as a healthy machine.

[0073] For example, when the training data is from an abnormal machine, if the machine is abnormal, its operating parameters are close to those of the training data, and the output evaluation value is small. The preset evaluation value threshold corresponds to the range of small evaluation values. When the training data is from a healthy machine, if the machine is abnormal, its operating parameters are significantly different from those of the training data, and the output evaluation value is large. The preset evaluation value threshold corresponds to the range of large evaluation values.

[0074] Therefore, in the machine anomaly detection method of this application, the processor 30 analyzes the processing results and working parameters through a loss function, and compares the loss value as an evaluation value with a preset evaluation value threshold to determine the machine status. The anomaly detection method can accurately analyze the machine status based on the working parameters to ensure that machines in an abnormal state can be accurately and quickly identified so that operators can maintain the machine in a timely manner to ensure the yield rate of the generated products.

[0075] Please refer to Figure 1 and Figure 6 In some implementations, 09: Determining whether a machine is abnormal based on the processing result includes: 093: When there is more than one machine, input the processing results and working parameters into the loss function simultaneously to output an evaluation value. Sort the evaluation values ​​of each machine in ascending or descending order. If the evaluation value is within the preset sequence threshold, the machine is determined to be an abnormal machine; if the evaluation value is outside the preset sequence threshold, the machine is determined to be a healthy machine.

[0076] Furthermore, please combine Figure 9 The processor 30 is also used to execute the method in 093. Specifically, the processor 30 is configured to: when there is more than one machine, simultaneously input the processing result and working parameters into the loss function to output an evaluation value. The evaluation values ​​of each machine are arranged in ascending or descending order; if the evaluation value is within a preset sequence threshold, the machine is determined to be an abnormal machine; if the evaluation value is outside the preset sequence threshold, the machine is determined to be a healthy machine.

[0077] Specifically, in the above embodiments, the concepts of machine tool abnormality and health are relative and determined by machining accuracy and machining quality. For example, when a machine tool is able to perform the machining process normally, if the machining quality requirements are high, a machine tool with a slight deviation in a certain working parameter will be identified as an abnormal machine tool; if the machining quality requirements are low, a machine tool with a large deviation in a certain working parameter may still be identified as a healthy machine tool.

[0078] Therefore, in the above embodiments, the determination of the machine status adopts a relative standard rather than an absolute standard. Upon obtaining the evaluation values ​​of each machine, the processor 30 sorts all the evaluation values ​​in ascending or descending order and determines the machine status based on a sequence threshold.

[0079] For example, when the training data consists of data from abnormal machines, if a machine is indeed abnormal, its operating parameters are close to those of the training data. In this case, the output evaluation value is smaller; that is, the smaller the evaluation value, the higher the degree of abnormality of the machine. When the evaluation values ​​of the machines are in ascending order, the machines with higher ranking numbers are considered more abnormal. In this case, a preset sequence threshold can be set to 0-20%, meaning that the top 20% of machines are considered abnormal. The same logic applies to descending order, and will not be elaborated further here.

[0080] When the training data consists of data from healthy machines, if a machine is considered abnormal, its operating parameters will differ significantly from the training data. In this case, the output evaluation value will be larger; a higher evaluation value indicates a higher degree of abnormality. When the evaluation values ​​of the machines are arranged in ascending order, machines ranked later in the list are considered more abnormal. Therefore, a preset sequence threshold of 80%-100% can be set, meaning that the bottom 20% of machines are considered abnormal. The same logic applies to descending order, and will not be elaborated further here.

[0081] Therefore, in the machine abnormality detection method of this application, the processor 30 analyzes the processing results and working parameters through a loss function and sorts the loss values ​​to determine that the machine located within the sequence threshold is an abnormal machine. At this time, the detection method can be flexibly adjusted according to the requirements of processing quality to ensure that the machine in an abnormal state can be accurately and quickly identified so that the operator can maintain the machine in time to ensure the yield rate of the generated products.

[0082] Please refer to Figure 1 and Figure 7 In some implementations, the anomaly detection method further includes: 051: Before inputting the working parameters into the trained detection model for processing, the anomaly detection method also includes: Historical abnormal parameters of multiple abnormal machine tools are acquired to form a training working parameter set. These historical abnormal parameters are obtained through at least one processing operation of the abnormal machine tool during its abnormal state. The historical abnormal parameters include at least one of the following: current value, voltage value, temperature, torque, speed, power, load, and feed rate of the abnormal machine tool during its abnormal operating state. 053: Input the training parameter set into the initial training model for training to obtain the trained detection model.

[0083] Furthermore, please combine Figure 9 The processor 30 is also used to execute the methods in 051 and 053. Specifically, the processor 30 is configured to: acquire historical abnormal parameters of multiple abnormal machines to form a training working parameter set, wherein the historical abnormal parameters are acquired through at least one processing operation of the abnormal machine in an abnormal state, and the historical abnormal parameters include at least one of the following: current value, voltage value, temperature, torque, speed, power, load, and feed rate of the abnormal machine in the working state; and input the training working parameter set into the initial training model for training to obtain a trained detection model.

[0084] Specifically, in the above implementation, the machine tool collects its operating parameters under abnormal conditions for training the detection model. These operating parameters under abnormal conditions are the historical anomaly parameters. The set of historical anomaly parameters is the training operating parameter set. Similar to the operating parameters mentioned above, historical anomaly parameters are specific parameter values ​​from at least one processing stage during which the machine tool was in an abnormal state. These parameter values ​​are collected from existing detection components required for the machine tool to perform the basic processing. In this case, the machine tool does not need to be equipped with additional detection components dedicated entirely to anomaly detection and not involved in the basic processing.

[0085] Similarly, historical abnormal parameters correspond to at least one of the following when the machine is in an abnormal state: current value, voltage value, temperature, torque, speed, power, load, and feed rate.

[0086] The initial training model is a pre-defined model for training. In this application, the initial training model includes a CNN+Transformer model. The processor 30 inputs the aforementioned training parameter set into the initial training model to train it and obtain a trained detection model. It can be understood that, at this point, if the machine is abnormal, its operating parameters are close to the training data, resulting in a smaller output evaluation value. A smaller evaluation value indicates a higher degree of abnormality of the machine, and the machine's state can be determined based on the evaluation.

[0087] Therefore, in the machine anomaly detection method of this application, the processor 30 uses historical anomaly parameters of the abnormal machine to form a training working parameter set, and uses the training working parameter set to train an initial training model to obtain a trained detection model. At this time, the evaluation value corresponding to the machine is relatively small when the machine is abnormal, so the processor 30 can accurately analyze the machine's state using the evaluation value. The parameters used in the acquisition of the trained detection model are of consistent types and have relatively rich data types, therefore the trained detection model has good performance.

[0088] Please refer to Figure 1 , Figure 7 and Figure 8 In some implementations, the initial training model includes a first module and a second module. The first module includes a CNN module, an upsampling module, and a dynamic weight acquisition module, and the second module includes an encoding module, a decoding module, and a fully connected module. The position vectors are configured to sort the inputs to the second processing layer other than the position vectors. 053: The training working parameter set is input into the initial training model for training to obtain a trained detection model, including: 0531: Input at least a portion of the historical outlier parameters from the training parameter set into the CNN module, the upsampling module, and the dynamic weight acquisition module respectively to obtain the first intermediate training parameters; 0532: The encoding module extracts features from the first intermediate training parameters after sorting the position vectors; 0532: The decoding module restores the feature extraction results of the encoding module; 0533: The fully connected module reshapes the reconstruction result of the decoding module so that the training result output by the second module has the same data form as the historical anomaly parameters; 0534: Input the training results and historical outlier parameters into the loss function to output a feedback value; 0535: If the feedback value is greater than the feedback value threshold, update the initial training model based on the feedback value; and 0536: Input at least a portion of the remaining data of the training working parameter set into the updated initial training model, and use the updated initial training model to perform the steps of obtaining the first intermediate training parameters to updating the initial training model according to the feedback value, until the feedback value is less than or equal to the feedback value threshold, then use the current initial training model as the detection model that has been trained.

[0089] Furthermore, please combine Figure 9 The processor 30 is also configured to execute the methods in 0531, 0532, 0533, 0534, 0535, and 0536. Specifically, the processor 30 is configured to: input at least a portion of the historical anomaly parameters from the training working parameter set into a CNN module to obtain first intermediate training parameters; perform feature extraction on the first intermediate training parameters after sorting the position vectors by an encoding module; restore the feature extraction results of the encoding module by a decoding module; shape the restored results of the decoding module by a fully connected module so that the training results output by the second module have the same data form as the historical anomaly parameters; input the training results and historical anomaly parameters into a loss function to output a feedback value; update the initial training model according to the feedback value if the feedback value is greater than a feedback value threshold; and input at least a portion of the remaining data from the training working parameter set into the updated initial training model, and use the updated initial training model to execute the steps from obtaining the first intermediate training parameters to updating the initial training model according to the feedback value, until the feedback value is less than or equal to the feedback value threshold, then use the current initial training model as the detection model that has been trained.

[0090] Specifically, the first module and the second module are the internal structures of the initial training model, which are distributed sequentially and work together to process historical anomaly parameters. At least some of the historical anomaly parameters are first fed into the first module after being input into the initial training model. The intermediate parameters obtained by the first module are then further input into the second module for analysis and processing, and the second module outputs the training results.

[0091] More specifically, in the first module, the CNN module is the core feature extractor in the initial trained model. The CNN module includes a residual function and can analyze and perceive historical anomalous parameters of the input through a series of learnable convolutional kernels. The upsampling module is the part of the first module used to implement the identity mapping. The dynamic weight acquisition module is the part of the first module used to determine the weights corresponding to the distributions of the CNN module and the upsampling module (i.e., the weights corresponding to the residual function and the identity mapping).

[0092] When at least some historical anomaly parameters are input into the initial training model, these parameters are then processed by the first module. The dynamic weight acquisition module analyzes and processes the working parameters using its built-in algorithm to output weights corresponding to the distributions of the CNN module and the upsampled module, based on the specific data conditions. The CNN module performs feature extraction and other analysis steps on the working parameters, obtaining analysis results that correspond to the anomaly states. These analysis results are determined by the residual function within the CNN module.

[0093] The upsampling module performs a constant mapping on this portion of historical outlier parameters, meaning the first module retains the original data content of these historical outlier parameters. Furthermore, the first module integrates the outputs of the CNN module, the upsampling module, and the dynamic weight acquisition module as the first intermediate training parameters. In other words, the first intermediate training parameters are the output of the first module given this portion of historical outlier parameters as input.

[0094] The position vector is used to sort the other inputs to the second module, excluding the position vector itself. When the first intermediate training parameters and the position vector are used as inputs together, the position vector sorts the first intermediate training parameters so that the data in the first intermediate training parameters are arranged in a predetermined order.

[0095] The encoding module is the part of the second module used for feature extraction from the input. The encoding module can analyze the first intermediate training parameters, which are sorted by position vectors and have a specific order of internal data, thereby extracting features and generating feature extraction results, which are then sent to the decoding module.

[0096] The decoding module is the part of the second module used to restore the data from the feature extraction results generated by the encoding module. The decoding module can decode and restore the feature extraction results generated by the encoding module, thereby generating the restored result and sending the restored result to the fully connected module.

[0097] The fully connected module is the part of the second module used to shape the reconstruction results generated by the decoding module. The fully connected module shapes the data format in the reconstruction results generated by the decoding module and outputs the shaped result as the training result. It can be understood that the output of the fully connected module is the output of the second module. The processed result output by the fully connected module, after shaping, has the same data format as the historical anomaly parameters initially input into the first module.

[0098] Furthermore, the training results with the same data format are input into the loss function along with historical anomaly parameters, and the resulting loss value is output as the feedback value. Here, the loss function measures the difference between the data obtained by the initially trained model and the anomaly data. The smaller the feedback value, the smaller the difference between the training results and the historical anomaly parameters. This means the initially trained model can better analyze and represent the data from the anomaly machines; that is, the initially trained model can stably output data with small differences and strong correlation to the input, given historical anomaly parameters as input.

[0099] The feedback threshold represents the degree of difference between the data obtained by the initially trained model and the abnormal data. When the feedback value exceeds the feedback threshold, the initially trained model fails to consistently output data with small differences and strong correlation to the input, even when the input consists of historical abnormal parameters. Therefore, the initial training model needs further adjustment. At this point, the processor 30 updates the model based on the feedback value, for example, by adjusting the internal structure and parameters of the initial training model.

[0100] After the initial training model is updated, the processor 30 uses the remaining data from the training parameter set that did not participate in the above training process as validation data, and inputs at least a portion of this remaining data into the updated initial training model, and executes methods 0531 to 0535. This process is repeated until the feedback value obtained after each execution is less than or equal to the feedback value threshold. At this point, the initial training model used in this processing can stably output data with small differences from the input and strong correlation when the input is historical abnormal parameters. Therefore, the current initial training model is used as the trained detection model for executing the machine's anomaly detection method.

[0101] It can be understood that the first module in the current initial training model is the first processing layer in the trained detection model, wherein: the CNN module in the current initial training model is the CNN layer in the trained detection model, the upsampling module in the current initial training model is the upsampling layer in the trained detection model, and the dynamic weight acquisition module in the current initial training model is the dynamic weight acquisition layer in the trained detection model; the second module in the current initial training model is the second processing layer in the trained detection model, wherein: the encoding module in the current initial training model is the encoding layer in the trained detection model, the decoding module in the current initial training model is the decoding layer in the trained detection model, and the fully connected module in the current initial training model is the fully connected layer in the trained detection model.

[0102] Therefore, in the machine anomaly detection method of this application, the initial training model structure used in the training process of the completed detection model is the same as the structure of the completed detection model mentioned above. The completed detection model, obtained through repeated training and model adjustment, can stably output data with small differences and strong correlation to the input, even when the input is historical anomaly parameters. Therefore, this application utilizes training parameters composed of historical anomaly parameters and iterates and adjusts the initial training model to obtain the completed detection model. The completed detection model can accurately process and analyze the input data to accurately detect the machine's status.

[0103] The technical features of the embodiments described above can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. Furthermore, other implementation methods can be derived from the above embodiments, allowing for structural and logical substitutions and changes without departing from the scope of this disclosure.

[0104] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for detecting machine malfunctions, characterized in that, include: Obtain the machine's operating parameters; The working parameters are input into a preset, trained detection model for processing, and the processing result is output. The processing result output by the preset, trained detection model is determined at least by the residual function, the identity mapping, the first dynamic weight related to the residual function, and the second dynamic weight related to the identity mapping. and Based on the processing results, determine whether the machine is an abnormal machine.

2. The anomaly detection method according to claim 1, characterized in that, Also includes: The operating parameters are preprocessed; The working parameters are input into the trained detection model for processing to output the processing results, including: The preprocessed working parameters are input into the trained detection model for further processing to output the processing results.

3. The anomaly detection method according to claim 2, characterized in that, The preprocessing includes at least one of the following: noise reduction, normalization, and data normalization.

4. The anomaly detection method according to claim 2, characterized in that, The trained detection model includes a first processing layer, a second processing layer, and a position vector. The first processing layer includes a CNN layer, an upsampling layer, and a dynamic weight acquisition layer. The position vector is configured to sort all inputs to the second processing layer except for the position vector. The step of inputting the working parameters into the trained detection model for processing to output a processing result, or the step of inputting the preprocessed working parameters into the trained detection model for processing to output a processing result, includes: The dynamic weight acquisition layer processes the working parameters to obtain the first dynamic weight and the second dynamic weight; The CNN layer processes the working parameters to obtain a first intermediate parameter, which is determined by the residual function in the CNN layer. The upsampling layer performs a constant mapping on the working parameters to obtain the second intermediate parameters; The third intermediate parameter is obtained based on the first intermediate parameter, the first dynamic weight, the second intermediate parameter, and the second dynamic weight; The second processing layer processes the third intermediate parameter and the position vector to obtain the processing result.

5. The anomaly detection method according to claim 4, characterized in that, The second processing layer includes an encoding layer, a decoding layer, and a fully connected layer; the second processing layer processes the third intermediate parameter and the position vector to obtain the processing result, including: The encoding layer extracts features from the third intermediate parameter after sorting the position vectors; The decoding layer restores the feature extraction results of the encoding layer to obtain the restored result; The fully connected layer shapes the restoration result so that the processing result output by the second processing layer has the same data format as the working parameters.

6. The anomaly detection method according to claim 1, characterized in that, The operating parameters are obtained from at least one processing process in the historical processing process of the machine tool, and the operating parameters include at least one of the following: current value, voltage value, temperature, torque, speed, power, load and feed rate of the machine tool in the working state.

7. The anomaly detection method according to claim 1, characterized in that, The step of determining whether the machine is abnormal based on the processing result includes: The processing result and the working parameters are simultaneously input into a loss function to output an evaluation value. If the evaluation value is within a preset evaluation value threshold, the machine is determined to be an abnormal machine; if the evaluation value is outside the preset evaluation value threshold, the machine is determined to be a healthy machine.

8. The anomaly detection method according to claim 1, characterized in that, Before inputting the working parameters into the trained detection model for processing, the method further includes: Historical abnormal parameters of multiple abnormal machine tools are acquired to form a training working parameter set. These historical abnormal parameters are obtained through at least one historical processing operation of the abnormal machine tool during its abnormal state. The historical abnormal parameters include at least one of the following: current value, voltage value, temperature, torque, speed, power, load, and feed rate of the abnormal machine tool during its abnormal operating state. The training parameter set is input into the initial training model for training to obtain the trained detection model.

9. The anomaly detection method according to claim 8, characterized in that, The initial training model includes a first module and a second module. The first module includes a CNN module, an upsampling module, and a dynamic weight acquisition module. The second module includes an encoding module, a decoding module, and a fully connected module. The position vector is configured to sort the inputs to the second module other than the position vector; The step of inputting the training parameter set into the initial training model for training to obtain the trained detection model includes: At least a portion of the historical anomaly parameters from the training parameter set are respectively input into the CNN module, the upsampling module, and the dynamic weight acquisition module to obtain the first intermediate training parameters; The encoding module extracts features from the first intermediate training parameters after sorting the position vectors; The decoding module restores the feature extraction results of the encoding module; The fully connected module reshapes the restoration result of the decoding module so that the training result output by the second module has the same data form as the historical anomaly parameters; The training results and the historical anomaly parameters are input into the loss function to output a feedback value; If the feedback value is greater than a feedback value threshold, the initial training model is updated based on the feedback value; and At least a portion of the remaining data of the training working parameter set is input into the updated initial training model, and the updated initial training model is used to perform the steps from obtaining the first intermediate training parameters to updating the initial training model according to the feedback value, until the feedback value is less than or equal to the feedback value threshold, and then the current initial training model is used as the detection model after training is completed.

10. An abnormal machine tool detection device, comprising a memory and a processor, wherein the memory is used to store instructions, characterized in that, The instructions stored in the memory are executed by the processor to implement claim 1. The anomaly detection method as described in any one of the 9.

Images