Method for predicting safe service life of equipment in manufacturing industry

Abstract

The invention provides a manufacturing industry equipment safety service life prediction method, and relates to the field of machine learning. The invention provides a CAMENet prediction model which comprises a composite attention module, an expert mixing module and an integration module, the composite attention module enables the model to master local and global time dependence at the same time and improve the expression ability of the model by fusing intra-segment attention and inter-segment attention, and meanwhile, a dynamic scaling factor is introduced, so that the prediction accuracy of the CAMENet prediction model is improved. The attention weight can be dynamically adjusted by the model according to the importance of each time step, so that the prediction performance is improved, the expert weight can be adaptively generated by the expert hybrid module according to the input equipment safety service life data characteristics, the optimal fragment division size is dynamically selected, the flexibility and generalization ability of the model are improved, and the prediction efficiency is improved. And the integration module performs weighted aggregation according to the expert weight to ensure that the model can adaptively integrate useful information in the multi-scale feature extraction process.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine learning, and in particular relates to a method for predicting the safe service life of manufacturing equipment. Background Art

[0002] Predicting the safe service life of manufacturing equipment is a key technology to ensure efficient factory operation and avoid unexpected downtime. Traditionally, equipment maintenance mainly relies on regular inspections and empirical judgments. This method is not only time-consuming and labor-intensive, but also difficult to detect potential equipment failure risks in advance. In addition, preventive maintenance strategies based on fixed time intervals may lead to unnecessary downtime and waste of resources, and fail to maximize the service life of equipment.

[0003] The safe service life data of manufacturing equipment is multidimensional time series data, which has complex nonlinear relationships between equipment failures and service life and various factors. Since the data is extracted from sensors, it may contain a lot of noise and missing values, making prediction difficult. Current machine learning has shown significant advantages in processing time series data and nonlinear relationships. It improves prediction accuracy through technical means such as feature extraction and model training, and is suitable for classification of manufacturing equipment status and prediction of equipment safe service life.

[0004] At present, there are several shortcomings in the technology for predicting the safe service life of equipment. Traditional machine learning methods are difficult to process massive sensor data and complex equipment operating conditions, resulting in poor accuracy and real-time performance of the prediction model. Existing machine learning technologies often only focus on a single parameter of the equipment, lack comprehensive analysis of multiple parameters, and cannot fully reflect the health status of the equipment. Most machine learning prediction models require a lot of adjustments and retraining when facing different types of equipment and working conditions, and have poor adaptability and universality. In response to the above problems, the present invention proposes a method for predicting the safe service life of manufacturing equipment. Summary of the invention

[0005] The present invention provides a method for predicting the safe service life of manufacturing equipment, aiming to propose a CAMENet prediction model, which consists of a composite attention module, an expert hybrid module and an integration module. The composite attention module can focus on local and global features at the same time to improve the expression ability of the model. The expert hybrid module can adaptively select a path according to the input equipment safe service life data characteristics to improve the flexibility and generalization ability of the model. The integration module improves the accuracy and stability of the prediction results by integrating multi-scale features.

[0006] The present invention proposes a safe service life prediction model for manufacturing equipment, which specifically includes the following steps: S1. Collect data related to the safe service life of manufacturing equipment, including characteristics and target variables, and pre-process the collected data; S2. Use batch normalization to normalize the preprocessed data and divide the data set; S3. Construct a composite attention module to process the local and global information of equipment safety service life data, which specifically includes the following steps: S31. Input the equipment safety service life time series data Divided into different sizes fragments; S32, introduce dynamic scaling factor in the fragment , perform attention calculation in each fragment and concatenate the attention results of all fragments; S33, introduce dynamic scaling factors between segments , perform attention calculation between segments, and weightedly fuse the attention results within and between segments; S4. Construct an expert hybrid module to process the periodicity and trend patterns of equipment safety service life data. The specific steps are as follows: S41. Perform period decomposition and use fast Fourier transform to convert the equipment safety service life time series data Convert to frequency domain and select the one with the largest amplitude The frequency components are inversely transformed to obtain the periodic quantity ; S42, perform trend decomposition and , using average pooling with different kernels for exponential moving average, we get the trend ; S43. Perform expert selection and generate expert weights using decomposition results , and select the value with the largest weight; S5. Construct an integration module to perform weighted aggregation on the outputs of the composite attention module and the expert mixture module to obtain the output ; S6. Use the fully connected layer to process features of different scales and obtain the predicted value of the safe service life of manufacturing equipment .

[0007] Preferably, in step S1, the data related to the safe service life of manufacturing equipment includes equipment operating time, equipment usage frequency, equipment working environment temperature, equipment load condition, equipment vibration data, equipment current and voltage sensor data, equipment maintenance history data, and equipment maintenance cycle data. The principle is to handle outliers in the safe service life data of manufacturing equipment. The specific formula is as follows: ; In the formula, is an outlier, Equipment safety service life data The mean of To prepare safe service life data The standard deviation of .

[0008] Preferably, in step S2, the preprocessed data is normalized using a layer normalization method, and the data set is divided proportionally. The specific formula is as follows: ; In the formula, It is the time series data of equipment safety service life. is the mean of the small batch, is the standard deviation of the small batch, is a learnable scaling parameter, is a learnable offset parameter.

[0009] Preferably, the steps S3 and S31 define The set of fragment size values ​​is , each fragment size A split operation corresponding to the fragment, which converts the equipment safety service life time series data Split into fragments, respectively , the specific formula is as follows: ; In the formula, is the length of the equipment safety service life time series data, is the fragment size, then each fragment is , is the characteristic dimension of the equipment's safe service life, For the The size of the fragment.

[0010] Preferably, fragmentation processing divides the long-term equipment safety service life time series data into multiple smaller fragments, so that the model can process the data features of each fragment more carefully, which helps the model capture the details of the data in a local range while retaining the overall time sequence information. Specifically, each fragment represents the operating status of the equipment in a certain time period, and multiple fragments are combined to reflect the overall usage of the equipment. It not only improves the expressiveness of the data, but also enables the model to process data of different lengths and time steps, thereby improving the flexibility and adaptability of the model. The division of fragments enables the model to better capture the details of the data in a local range and improves the accuracy of feature extraction.

[0011] Preferably, in steps S3 and S32, Clips , first along the feature dimension Perform embedding transformation to convert it into a new feature embedding dimension , get the representation within the fragment , obtain the query by linear transformation ,key Sum , and introduce a dynamic scaling factor , to calculate the attention within the fragment, the specific formula is as follows: ; In the formula, , , are the query matrix, key matrix, and value matrix within the fragment, is the feature embedding dimension after embedding, is the dynamic scaling factor within the segment, and the attention weight can be dynamically adjusted according to the input data. After the attention calculation within the segment, The size of the fragment is determined by Transformed into , concatenate the attention results in all segments, the specific formula is as follows: ; In the formula, It is the concatenated output of the attention mechanism results within all fragments, with a dimension of , For the Clips The internal attention results of the fragment, For splicing operation.

[0012] Preferably, the dynamic scaling factor enables the model to dynamically adjust the attention weight according to the importance of each time step, thereby improving the accuracy of feature representation and enhancing the model's ability to recognize the importance of different time steps, making the model more flexible and accurate in processing local features. In each fragment, the query, key, and value are obtained through linear transformation, and attention calculation is performed. The attention results of each fragment are connected, which can effectively capture local features. The feature information of each fragment is integrated through the connection operation, further improving the effect of feature extraction.

[0013] Preferably, in the steps S3 and S33, the equipment safety service life time series data divided into segments is Along the feature dimension Embedded into feature embedding dimension , and rearrange the embedded data to reduce the fragment size and feature embedding Combine to get the representation between fragments ,in , obtain the query by linear transformation ,key Sum , and introduce a dynamic scaling factor , to calculate the attention between fragments, the specific formula is as follows: ; In the formula, , , are the query matrix, key matrix and value matrix between fragments, is the feature combination dimension after embedding, is a dynamic scaling factor between fragments. Combined with the adaptive weight strategy, the representation within the fragment is transformed through a linear layer. and representation between fragments Time series data on equipment safety service life The combination generates adaptive weights, the specific formula is as follows: ; ; In the formula, It is the time series data of equipment safety service life. is the representation within the fragment, is the representation between fragments, is the input data of adaptive weight, is the learnable weight matrix, is the bias term, and the attention results within and between segments are weighted and fused through adaptive weights. The specific formula is as follows: ; In the formula, , is the adaptive weight, Attention results within a fragment, is the attention result between fragments.

[0014] Preferably, the dynamic scaling factor between fragments enables the model to dynamically adjust the attention weights between fragments, thereby enhancing the ability to capture global features and improving the flexibility and accuracy of the model in processing global features. In the attention calculation between fragments, the query, key and value are obtained through linear transformation. Combined with the adaptive weight strategy, the attention results within and between fragments are weighted and fused, which can comprehensively consider the features at different scales and enhance the comprehensive understanding of the data.

[0015] Preferably, in the steps S4 and S41, the equipment safety service life time series data is converted into Decompose into frequency domain and calculate the amplitude of each frequency component in the frequency domain. The specific formula is as follows: ; ; In the formula, It is the time series data of equipment safety service life. is the fast Fourier transform, is the representation of the equipment safety service life time series data in the frequency domain, Time series data for equipment safety service life In the frequency domain The complex value of the frequency components, , is the frequency component The real and imaginary parts of the The index of the frequency component is as follows: ; ; In the formula, is the index array obtained by sorting the amplitudes of the frequency components in ascending order, To select the largest value in the ascending sorted index The indices of the frequency components are used to construct a frequency domain mask, retaining the most significant frequency components, and perform inverse fast Fourier transform on the masked frequency domain representation to obtain the periodicity. The specific formula is as follows: ; ; In the formula, Time series data for equipment safety service life In the frequency domain The complex value of the frequency components, is the inverse fast Fourier transform, is the frequency domain representation after masking.

[0016] Preferably, the fast Fourier transform is used to convert the equipment's safe service life time series data into the frequency domain, and the frequency component with the largest amplitude is selected for inverse transformation to obtain the periodic quantity. The fast Fourier transform can convert the time series data into a frequency domain representation, thereby identifying the periodic components in the data, enabling the model to accurately capture the periodic changes in equipment operation, enhancing the understanding and prediction capabilities of the data periodicity, and this type of periodic decomposition method helps to reveal the periodic patterns in the data, providing a basis for subsequent trend analysis.

[0017] Preferably, in the steps S4 and S42, the original equipment safety service life time series data is obtained. Subtract the periodic amount from the result to get the residual part. The specific formula is as follows: ; In the formula, It is the time series data of equipment safety service life. For the periodic quantity of equipment safety service life data, use different verifications for the remaining parts , perform exponential moving average operation, for each core , perform exponential moving average operation and average pooling operation, the specific formula is as follows: ; ; In the formula, is the exponential moving average function, To use The result of exponential moving average of the cores is: is the average pooling function, To use The result of average pooling of cores is For the remaining part, weighted operations are performed on the moving average results and average pooling results of different cores. The specific formula is as follows: ; In the formula, is the linear transformation function, is the number of cores, For the remaining part, To use The result of average pooling of cores is To use The result of exponential moving average of the cores.

[0018] Preferably, the remaining part is subjected to exponential moving average using average pooling with different kernels to obtain the trend quantity. The trend decomposition method extracts the long-term trend components in the data and combines average pooling with different kernels and smoothing operations to effectively remove noise in the data, extract trend characteristics of the data, and improve the ability to capture data trends. Such methods can help the model perform better in long-term predictions.

[0019] Preferably, in steps S4 and S43, the original equipment safety service life time series data is The number of cycles in which equipment safety service life data is added Trend quantity of equipment safety service life data Perform linear transformation and transform and merge along the time dimension. The specific formula is as follows: ; In the formula, is a linear mapping function, It is the time series data of equipment safety service life. is the periodic quantity of equipment safety service life data, is the trend value of the equipment safety service life data, and then the expert function is used to generate the expert weight, and the value with the largest weight is selected. The specific formula is as follows: ; ; In the formula, The results of cycle and trend decomposition are: , The learnable parameters generated for the weights, is the normally distributed noise term, is a smoothed version of the ReLU function, used to ensure the non-negativity of the output. The expert weights generated for the expert function.

[0020] Preferably, the decomposition results are used to generate expert weights, and the value with the largest weight is selected. The expert selection mechanism generates expert weights according to the results of cycle and trend decomposition, and dynamically selects the most suitable fragment division size to optimize the feature representation. It can adaptively select the optimal feature representation, improve the flexibility and generalization ability of the model, and by selecting the value with the largest weight, the model can dynamically adjust its own structure to adapt to different data characteristics, improve prediction accuracy, and at the same time improve the adaptability and flexibility of the model.

[0021] Preferably, in step S5, the outputs of the composite attention module and the expert mixture module are weighted aggregated, and the specific formula is as follows: ; In the formula, Different fragment sizes selected for expert weights The attention results, is the feature alignment operation, is an indicator function that outputs 1 when the expert weight is greater than 0, otherwise it outputs 0. Decompose the results into cycles and trends.

[0022] Preferably, the outputs of the composite attention module and the expert mixture module are weighted aggregated. Weighted aggregation ensures that the model can adaptively integrate useful information in the multi-scale feature extraction process by comprehensively utilizing the outputs of different modules, enhances the model's ability to integrate multi-scale features, and improves the accuracy and stability of the prediction results. By weighted processing of the outputs of different modules, the model can more comprehensively capture the features in the data and improve the overall understanding of the data and the prediction effect.

[0023] Preferably, in step S6, the output after weighted aggregation is processed by a fully connected layer. Processing is performed to obtain the predicted value of the safe service life of manufacturing equipment. The specific formula is as follows: ; In the formula, is the output after weighted aggregation, is the weight matrix of the fully connected layer, is the bias vector of the fully connected layer, It is the predicted value of the safe service life of the equipment.

[0024] In summary, due to the adoption of the technical solution, the beneficial effects of the present invention are as follows: the present invention proposes a CAMENet prediction model, which is applied to the safe service life prediction scenario of manufacturing equipment, and is composed of a composite attention module, an expert hybrid module and an integration module. The composite attention module enables the model to simultaneously grasp local and global time dependencies by fusing intra-segment attention and inter-segment attention, thereby improving the expressiveness of the model. At the same time, a dynamic scaling factor is introduced, so that the model can dynamically adjust the attention weight according to the importance of each time step, thereby improving the prediction performance. The expert hybrid module can adaptively generate expert weights according to the input equipment safe service life data characteristics, dynamically select the optimal segment division size, and improve the flexibility and generalization ability of the model. The integration module performs weighted aggregation according to the expert weights to ensure that the model can adaptively integrate useful information in the multi-scale feature extraction process. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a step-by-step diagram of a method for predicting the safe service life of manufacturing equipment.

[0026] Figure 2 This is the structure diagram of the CAMENet prediction model.

[0027] Figure 3 This is the structural diagram of the composite attention module.

[0028] Figure 4 This is the structure diagram of the expert hybrid module.

[0029] Figure 5This is the fitting effect diagram of the CAMENet prediction model for predicting the safe service life of manufacturing equipment. DETAILED DESCRIPTION

[0030] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0031] See also Figure 1-Figure 5 The present invention provides a technical solution: a method for predicting the safe service life of manufacturing equipment, which can focus on local and global features at the same time by constructing a composite attention module to improve the expression ability of the model, and the expert hybrid module can adaptively select paths according to the input equipment safety service life data characteristics to improve the flexibility and generalization ability of the model. The integration module improves the accuracy and stability of the prediction results by integrating multi-scale features. The specific steps are as follows: Figure 1 shown.

[0032] Construct the CAMENet prediction model, whose structure is as follows Figure 2 As shown, the specific steps are as follows: S1. Collect data related to the safe service life of manufacturing equipment, including characteristics and target variables, and preprocess the collected data.

[0033] Furthermore, the data related to the safe service life of manufacturing equipment includes equipment operating time, equipment usage frequency, equipment working environment temperature, equipment load condition, equipment vibration data, equipment current and voltage sensor data, equipment maintenance history data, equipment maintenance cycle data, and equipment maintenance cycle data. The principle is to handle outliers in the safe service life data of manufacturing equipment. The specific formula is as follows: ; In the formula, is an outlier, Equipment safety service life data The mean of To prepare safe service life data The standard deviation of .

[0034] S2. Use batch normalization to normalize the preprocessed data and divide the data set into two parts in a ratio of 8:2.

[0035] Furthermore, the preprocessed data is normalized using the layer normalization method, and the data set is divided proportionally. The specific formula is as follows: ; In the formula, It is the time series data of equipment safety service life. is the mean of the small batch, is the standard deviation of the small batch, is a learnable scaling parameter, is a learnable offset parameter.

[0036] S31. Input the equipment safety service life time series data Divided into different sizes fragments.

[0037] Furthermore, a composite attention module is constructed, whose structure is as follows Figure 3 As shown, the definition The set of fragment size values ​​is , each fragment size A split operation corresponding to the fragment, which converts the equipment safety service life time series data Split into fragments, respectively , the specific formula is as follows: ; In the formula, is the length of the equipment safety service life time series data, is the fragment size, then each fragment is , is the characteristic dimension of the equipment's safe service life, For the The size of the fragment.

[0038] S32, introduce dynamic scaling factor in the fragment , perform attention calculation in each fragment and concatenate the attention results of all fragments.

[0039] Furthermore, for Clips , first along the feature dimension Perform embedding transformation to convert it into a new feature embedding dimension , get the representation within the fragment , obtain the query by linear transformation ,key Sum , and introduce a dynamic scaling factor , to calculate the attention within the fragment, the specific formula is as follows: ; In the formula, , , are the query matrix, key matrix, and value matrix within the fragment, is the feature embedding dimension after embedding, is the dynamic scaling factor within the segment, and the attention weight can be dynamically adjusted according to the input data. After the attention calculation within the segment, The size of the fragment is determined by Transformed into , concatenate the attention results in all segments, the specific formula is as follows: ; In the formula, It is the concatenated output of the attention mechanism results within all fragments, with a dimension of , For the Clips The internal attention results of the fragment, For splicing operation.

[0040] S33, introduce dynamic scaling factors between segments , perform attention calculation between fragments, and weightedly fuse the attention results within and between fragments.

[0041] Furthermore, the equipment safety service life time series data divided into segments Along the feature dimension Embedded into feature embedding dimension , and rearrange the embedded data to reduce the fragment size and feature embedding Combine to get the representation between fragments ,in , obtain the query by linear transformation ,key Sum , and introduce a dynamic scaling factor , to calculate the attention between fragments, the specific formula is as follows: ; In the formula, , , are the query matrix, key matrix and value matrix between fragments, is the feature combination dimension after embedding, is a dynamic scaling factor between fragments. Combined with the adaptive weight strategy, the representation within the fragment is transformed through a linear layer. and representation between fragments Time series data on equipment safety service life The combination generates adaptive weights, the specific formula is as follows: ; ; In the formula, It is the time series data of equipment safety service life. is the representation within the fragment, is the representation between fragments, is the input data of adaptive weight, is the learnable weight matrix, is the bias term, and the attention results within and between segments are weighted and fused through adaptive weights. The specific formula is as follows: ; In the formula, , is the adaptive weight, Attention results within a fragment, is the attention result between fragments.

[0042] S41. Perform period decomposition and use fast Fourier transform to convert the equipment safety service life time series data Convert to frequency domain and select the one with the largest amplitude The frequency components are inversely transformed to obtain the periodic quantity .

[0043] Furthermore, we construct an expert mixture module, whose structure is as follows: Figure 4 As shown in the figure, the fast Fourier transform is used to transform the equipment safety service life time series data Decompose into frequency domain and calculate the amplitude of each frequency component in the frequency domain. The specific formula is as follows: ; ; In the formula, It is the time series data of equipment safety service life. is the fast Fourier transform, is the representation of the equipment safety service life time series data in the frequency domain, Time series data for equipment safety service life In the frequency domain The complex value of the frequency components, , is the frequency component The real and imaginary parts of the The index of the frequency component is as follows: ; ; In the formula, is the index array obtained by sorting the amplitudes of the frequency components in ascending order, To select the largest value in the ascending sorted index The indices of the frequency components are used to construct a frequency domain mask, retaining the most significant frequency components, and perform inverse fast Fourier transform on the masked frequency domain representation to obtain the periodicity. The specific formula is as follows: ; ; In the formula, Time series data for equipment safety service life In the frequency domain The complex value of the frequency components, is the inverse fast Fourier transform, is the frequency domain representation after masking.

[0044] S42, perform trend decomposition and , using average pooling with different kernels for exponential moving average, we get the trend .

[0045] Furthermore, from the original equipment safety service life time series data Subtract the periodic amount from the result to get the residual part. The specific formula is as follows: ; In the formula, It is the time series data of equipment safety service life. For the periodic quantity of equipment safety service life data, use different verifications for the remaining parts , perform exponential moving average operation, for each core , perform exponential moving average operation and average pooling operation, the specific formula is as follows: ; ; In the formula, is the exponential moving average function, To use The result of exponential moving average of the cores is: is the average pooling function, To use The result of average pooling of cores is For the remaining part, weighted operations are performed on the moving average results and average pooling results of different cores. The specific formula is as follows: ; In the formula, is the linear transformation function, is the number of cores, For the remaining part, To use The result of average pooling of cores is To use The result of exponential moving average of the cores.

[0046] S43. Perform expert selection and generate expert weights using decomposition results , and select the value with the largest weight.

[0047] Furthermore, the original equipment safety service life time series data The number of cycles in which equipment safety service life data is added Trend quantity of equipment safety service life data Perform linear transformation and transform and merge along the time dimension. The specific formula is as follows: ; In the formula, is a linear mapping function, It is the time series data of equipment safety service life. is the periodic quantity of equipment safety service life data, is the trend value of the equipment safety service life data, and then the expert function is used to generate the expert weight, and the value with the largest weight is selected. The specific formula is as follows: ; ; In the formula, The results of cycle and trend decomposition are: , The learnable parameters generated for the weights, is the normally distributed noise term, is a smoothed version of the ReLU function, used to ensure the non-negativity of the output. The expert weights generated for the expert function.

[0048] S5. Construct an integration module to perform weighted aggregation on the outputs of the composite attention module and the expert mixture module to obtain the output .

[0049] Furthermore, the outputs of the composite attention module and the expert mixture module are weighted and aggregated. The specific formula is as follows: ; In the formula, Different fragment sizes selected for expert weights The attention results, is the feature alignment operation, is an indicator function that outputs 1 when the expert weight is greater than 0, otherwise it outputs 0. Decompose the results into cycles and trends.

[0050] S6. Use the fully connected layer to process features of different scales and obtain the predicted value of the safe service life of manufacturing equipment .

[0051] Furthermore, the weighted aggregated output is processed through a fully connected layer. Processing is performed to obtain the predicted value of the safe service life of manufacturing equipment. The specific formula is as follows: ; In the formula, is the output after weighted aggregation, is the weight matrix of the fully connected layer, is the bias vector of the fully connected layer, It is the predicted value of the safe service life of the equipment.

[0052] Furthermore, the CAMENet prediction model is written in Python, using Pytorch as the framework in the CUDA11.27 environment, and trained on an NVIDIA A800 80GB GPU. The dataset is 30 years of safe service life data of manufacturing equipment, which is normalized and input into the CAMENet prediction model. The training batch is set to 128, and the training is performed using Optimizer, with the learning rate set to .

[0053] Furthermore, the CAMENet prediction model can achieve the prediction effect of safe service life of manufacturing equipment. Figure 5 As shown in the figure, the horizontal axis represents the equipment operation time (years), the vertical axis represents the remaining service life of the equipment (years), the black solid line represents the actual value, and the gray dotted line represents the predicted value. It can be seen from the figure that the predicted value of the CAMENet model is basically consistent with the actual value in most time periods, especially in the early stage (0-10 years) and mid-stage (10-20 years) of the equipment service life. The model prediction accuracy is high, and the prediction curve closely follows the actual curve, indicating that the model can accurately capture the changing trend of equipment life. During the entire equipment service life prediction process, the trend lines of the predicted value and the actual value are consistent, indicating that the model has good stability and robustness, especially when the equipment operation time is long (20-30 years). Despite the increase in data volatility, the model can still maintain the accuracy of the prediction well, indicating that it is stable when dealing with long-term prediction tasks. From the above analysis, it can be seen that the CAMENet prediction model performs well in the prediction of the safe service life of manufacturing equipment.

Claims

1. A method for predicting the safe service life of manufacturing equipment, characterized in that: The following steps are involved: S1. Collect data related to the safe service life of manufacturing equipment. Preprocessing of collected data including features and target variables; S2. Use batch normalization to normalize the preprocessed data and divide the data set; S3. Construct a composite attention module to process the local and global information of equipment safety service life data, which specifically includes the following steps: S31. Input the equipment safety service life time series data Divided into different sizes fragments; S32, introduce dynamic scaling factor in the fragment , perform attention calculation in each fragment and concatenate the attention results of all fragments; S33, introduce dynamic scaling factors between segments , perform attention calculation between segments, and weightedly fuse the attention results within and between segments; S4. Construct an expert hybrid module to process the periodicity and trend patterns of equipment safety service life data. The specific steps are as follows: S41. Perform period decomposition and use fast Fourier transform to convert the equipment safety service life time series data Convert to frequency domain and select the one with the largest amplitude The frequency components are inversely transformed to obtain the periodic quantity ; S42, perform trend decomposition and , using average pooling with different kernels for exponential moving average, we get the trend ; S43. Perform expert selection and generate expert weights using decomposition results , and select the value with the largest weight; S5. Construct an integration module to perform weighted aggregation on the outputs of the composite attention module and the expert mixture module to obtain the output ; S6. Use the fully connected layer to process features of different scales and obtain the predicted value of the safe service life of manufacturing equipment .

2. A method for predicting safe service life of manufacturing equipment according to claim 1, characterized in that: The steps S3 and S31 define The set of fragment size values ​​is , each fragment size A split operation corresponding to the fragment, which converts the equipment safety service life time series data Split into fragments, respectively , the specific formula is as follows: ； In the formula, is the length of the equipment safety service life time series data, is the fragment size, then each fragment is , is the characteristic dimension of the safe service life of the equipment, For the The size of the fragment.

3. A method for predicting safe service life of manufacturing equipment according to claim 2, characterized in that: In the steps S3 and S32, Clips , first along the feature dimension Perform embedding transformation to convert it into a new feature embedding dimension , get the representation within the fragment , obtain the query through linear transformation ,key Sum , and introduce a dynamic scaling factor , to calculate the attention within the fragment, the specific formula is as follows: ； In the formula, , , are the query matrix, key matrix, and value matrix within the fragment, is the feature embedding dimension after embedding, is the dynamic scaling factor within the segment, and the attention weight can be dynamically adjusted according to the input data. After the attention calculation within the segment, The size of the fragment is determined by Transformed into , concatenate the attention results in all segments, the specific formula is as follows: ； In the formula, It is the concatenated output of the attention mechanism results within all fragments, with a dimension of , For the Clips The internal attention results of the fragment, For splicing operation.

4. A method for predicting safe service life of manufacturing equipment according to claim 3, characterized in that: In the steps S3 and S33, the equipment safety service life time series data divided into segments is Along the feature dimension Embedded into feature embedding dimension , and rearrange the embedded data to reduce the fragment size and feature embedding Combine to get the representation between fragments ,in , obtain the query through linear transformation ,key Sum , and introduce a dynamic scaling factor , to calculate the attention between fragments, the specific formula is as follows: ； In the formula, , , are the query matrix, key matrix and value matrix between fragments, is the feature combination dimension after embedding, is a dynamic scaling factor between fragments. Combined with the adaptive weight strategy, the representation within the fragment is transformed through a linear layer. and representation between fragments Time series data on equipment safety service life The combination generates adaptive weights, the specific formula is as follows: ； ； In the formula, It is the time series data of equipment safety service life. is the representation within the fragment, is the representation between fragments, is the input data of adaptive weight, is the learnable weight matrix, is the bias term, and the attention results within and between segments are weighted and fused through adaptive weights. The specific formula is as follows: ； In the formula, , is the adaptive weight, Attention results within a fragment, is the attention result between fragments.

5. A method for predicting safe service life of manufacturing equipment according to claim 4, characterized in that: In the steps S4 and S41, the time series data of the equipment safety service life is converted into the time series data of the equipment safety service life by using the fast Fourier transform. Decompose into frequency domain and calculate the amplitude of each frequency component in the frequency domain. The specific formula is as follows: ； ； In the formula, It is the time series data of equipment safety service life. is the fast Fourier transform, is the representation of the equipment safety service life time series data in the frequency domain, Time series data for equipment safety service life In the frequency domain The complex value of the frequency components, , is the frequency component The real and imaginary parts of the The index of the frequency component is as follows: ； ； In the formula, is the index array obtained by sorting the amplitudes of the frequency components in ascending order, To select the largest value in the ascending sorted index The indices of the frequency components are used to construct a frequency domain mask, retaining the most significant frequency components, and perform inverse fast Fourier transform on the masked frequency domain representation to obtain the periodicity. The specific formula is as follows: ； ； In the formula, Time series data for equipment safety service life In the frequency domain The complex value of the frequency components, is the inverse fast Fourier transform, is the frequency domain representation after masking.

6. A method for predicting safe service life of manufacturing equipment according to claim 5, characterized in that: In the steps S4 and S42, the original equipment safety service life time series data is obtained Subtract the periodic amount from the result to get the residual part. The specific formula is as follows: ； In the formula, It is the time series data of equipment safety service life. For the periodic quantity of equipment safety service life data, use different verifications for the remaining parts , perform exponential moving average operation, for each core , perform exponential moving average operation and average pooling operation, the specific formula is as follows: ； ； In the formula, is the exponential moving average function, To use The result of exponential moving average of the cores is: is the average pooling function, To use The result of average pooling of cores is For the remaining part, weighted operations are performed on the moving average results and average pooling results of different cores. The specific formula is as follows: ； In the formula, is the linear transformation function, is the number of cores, For the remaining part, To use The result of average pooling of cores is To use The result of exponential moving average of the kernels.

7. A method for predicting safe service life of manufacturing equipment according to claim 6, characterized in that: In the steps S4 and S43, the original equipment safety service life time series data is The number of cycles in which equipment safety service life data is added and trend quantity of equipment safety service life data Perform linear transformation and transform and merge along the time dimension. The specific formula is as follows: ； In the formula, is a linear mapping function, It is the time series data of equipment safety service life. is the periodic quantity of equipment safety service life data, is the trend value of the equipment safety service life data, and then the expert function is used to generate the expert weight, and the value with the largest weight is selected. The specific formula is as follows: ； ； In the formula, The results of cycle and trend decomposition are: , The learnable parameters generated for the weights, is the normally distributed noise term, is a smoothed version of the ReLU function, used to ensure the non-negativity of the output. The expert weights generated for the expert function.

8. A method for predicting safe service life of manufacturing equipment according to claim 7, characterized in that: In step S5, the outputs of the composite attention module and the expert mixture module are weighted and aggregated, and the specific formula is as follows: ； In the formula, Different fragment sizes selected for expert weights The attention results, is the feature alignment operation, is an indicator function that outputs 1 when the expert weight is greater than 0, otherwise it outputs 0. Decompose the results into cycles and trends.

Images