Fault prediction method for door and window fitting equipment under high-load operation

Abstract

The invention provides a door and window fitting equipment fault prediction method under high-load operation, and relates to the field of door and window fitting equipment fault prediction. The invention provides a door and window fitting equipment fault prediction process under high-load operation. The method specifically comprises the steps of collecting data related to equipment operation states and faults, performing normalization processing on the collected data, providing a door and window fitting equipment fault prediction model under high-load operation, and training and testing the door and window fitting equipment fault prediction model under the high-load operation. The fault prediction model is composed of a comprehensive preprocessing module, a fusion attention mechanism, a coordination unit module and a layer-by-layer excitation network, and the comprehensive preprocessing module carries out blocking, embedding and coding processing on data; fusing an attention mechanism to establish a complex dependency relationship between a time dimension and a variable dimension; the coordination unit module introduces a coordination unit to aggregate and distribute attention information, so that the complexity is reduced; the layer-by-layer excitation network performs nonlinear transformation on the embedded representation of each data piece.

Description

Technical Field

[0001] The invention belongs to the field of door and window cutting equipment fault prediction, and in particular relates to a door and window cutting equipment fault prediction method under high-load operation. Background Art

[0002] With the rapid development of the construction industry, doors and windows, as an important part of buildings, have a growing production demand. As a key machine in the door and window manufacturing process, the operating status of door and window cutting equipment directly affects product quality and production efficiency. Under high-load operating conditions, the equipment is in a high-intensity operating state for a long time, which is prone to problems such as wear, overheating, and abnormal vibration, resulting in frequent equipment failures. This will not only cause production interruptions and increase maintenance costs, but may also affect delivery time and cause economic losses to the company.

[0003] Traditional equipment maintenance methods mainly include regular maintenance and post-maintenance. Although regular maintenance can prevent failures to a certain extent, there are often problems of over-maintenance or under-maintenance, and it is impossible to accurately maintain the actual status of the equipment. Post-maintenance is to repair the equipment after the failure occurs, which may cause production stagnation and affect production plans. With the continuous improvement of modern manufacturing industry's requirements for production efficiency and equipment reliability, traditional maintenance methods can no longer meet the needs of efficient production.

[0004] The emergence of fault prediction technology provides a new way to solve the above problems. By installing various sensors on the equipment, the key operating parameters of the equipment, such as vibration, temperature, current, voltage, etc., are monitored in real time, and a large amount of operating data is collected. By using technologies such as signal processing, pattern recognition and machine learning, the collected data is analyzed and modeled, which can predict possible equipment failures in advance and guide maintenance personnel to take preventive measures before failures occur to avoid production interruptions.

[0005] However, under high-load operating conditions, the operating environment of door and window cutting equipment is more complex, the failure modes are diverse, and traditional fault prediction methods may be insufficient in terms of accuracy and real-time performance. Therefore, it is necessary to study fault prediction methods suitable for high-load conditions. Combining advanced sensor technology, big data analysis and artificial intelligence algorithms can improve the accuracy and reliability of fault prediction. For example, deep learning algorithms are used to extract features and recognize patterns from massive data, monitor equipment status in real time, and detect abnormalities in a timely manner. Summary of the invention

[0006] The main purpose of the present invention is to provide a method for predicting the failure of door and window fitting equipment under high-load operation, aiming to construct a door and window fitting equipment failure prediction model under high-load operation, wherein the door and window fitting equipment failure prediction model under high-load operation is composed of a comprehensive preprocessing module, a fusion attention mechanism, a coordination unit module and a layer-by-layer excitation network, wherein the comprehensive preprocessing module divides, embeds and encodes the data; the fusion attention mechanism establishes a complex dependency relationship between the time dimension and the variable dimension; the coordination unit module aggregates and distributes attention information by introducing coordination units to reduce complexity; the layer-by-layer excitation network performs a nonlinear transformation on the embedded representation of each data slice.

[0007] To achieve the above object, the technical solution of the present invention is: a method for predicting failure of door and window fitting equipment under high load operation, the method comprising: S1. Collect data related to equipment operation status and faults, and monitor vibration, temperature, current, voltage and sound parameters in real time; S2. Normalize the collected data to eliminate the influence between different dimensions and scales; S3. A failure prediction model for door and window fitting equipment under high load operation is proposed. The specific method includes: S31, constructing a comprehensive preprocessing module to perform data segmentation, embedding and encoding processing; S32, build a fusion attention mechanism to establish complex dependencies between the time dimension and the variable dimension; S33, build a coordination unit module, by introducing coordination units to aggregate and distribute attention information to reduce complexity; S34, construct a layer-by-layer excitation network to perform nonlinear transformation on the embedded representation of each data slice; S4, training a door and window fitting equipment failure prediction model under high-load operation, wherein the training includes hyperparameter setting and model training process; S5. Testing the door and window fitting equipment failure prediction model under high load operation. The purpose of the test is to test whether the failure prediction model meets the requirements of the actual environment.

[0008] Furthermore, in the steps S3 and S31, the time series of each variable is divided into blocks according to the window length l to obtain multiple blocks. i ∈R T It is divided into p blocks of length l, and the formula is expressed as: In the formula, is the number of blocks, each block is the jth time period; Each block is transformed into an embedded representation through linear projection. In the embedded representation, each block is first flattened and converted into a one-dimensional vector. For each block of each vector, a linear transformation W∈R is performed. l×d Embedding it into d-dimensional space, the linear transformation formula is as follows: In the formula, is the embedding representation of the jth block, b is the bias term; after the entire data set is embedded, the embedding matrix H∈R N×p×d Block embedding for all variables; Then, position encoding is introduced to make the model have sequence order information. For the embedding of each time step, position encoding P∈R is added N×p×d , in order to retain the time order information, the added position encoding formula is: H′=H+P; Where H′ is data containing both variable and time position information.

[0009] Furthermore, in steps S3 and S31, blocking and position coding play an important role in the prediction of door and window fitting equipment failures under high-load operations. Blocking unifies the representation of different types of sensor data, enriches the feature expression capability, and effectively filters noise and enhances the correlation between multiple variables. Position coding retains time sequence information, enabling the model to identify trends and cycles in time and capture the time dependence of equipment failures, thereby improving prediction accuracy. The combination of the two enables the model to have stronger fault prediction capabilities in complex working environments.

[0010] Furthermore, in the steps S3 and S32, the fusion attention mechanism establishes a complex dependency relationship between the time dimension and the variable dimension. First, the input vector H′ is flattened to convert it into a one-dimensional sequence, and then a new vector X′∈R is obtained after expansion. N×p×d In the fusion attention mechanism, the query matrix Q, key matrix K and value matrix V are defined to calculate the attention. In order to capture the dependencies across time and variable dimensions at the same time, the fusion attention mechanism linearly transforms the query, key and value of each block. The formula is as follows: V=X′W V ∈R N×p×d ; Where W Q , W K , W V is the trainable linear transformation matrix, N is the number of variables, p is the number of blocks for each variable, and d kis the dimension of the embedding space, which is equal to the mapping dimension d; then the output of the fusion attention mechanism is calculated for each block, and the formula is as follows: In the formula, QK T To calculate the similarity between the query and the key, Scaling is used to stabilize the gradient, and the softmax operation is used to normalize the attention weights at each position. The matrix It is the information after the interaction between the time dimension and the variable dimension.

[0011] Furthermore, in the steps S3 and S32, the fusion attention mechanism has significant advantages in equipment failure prediction under high-load operation. The fusion attention mechanism can simultaneously pay attention to the time dimension and the dependency between each sensor variable at the same stage, so that the model can more accurately identify the complex interactions between multiple variables; in addition, the fusion attention mechanism can adapt to the dynamic changes between variables and flexibly capture the mutual influence of different sensor information under high-load conditions of equipment; by globally integrating information, the fusion attention mechanism helps the model to deeply identify fault modes and effectively optimize computing efficiency, thereby achieving higher accuracy and efficiency in fault prediction tasks.

[0012] Furthermore, in the steps S3 and S33, in the aggregation phase, the coordination unit performs cross attention calculation with the input O to aggregate the information in the input into the coordination unit; first, using D as the request matrix and O as the index and content matrix, the cross attention is calculated: the coordination unit request matrix is: Where k is the number of coordination units, W Q1 is the weight matrix; The input index matrix is: Where W K1 is the weight matrix; The input content matrix is: V O =OW V1 ∈R N×p×d Where W V1 is the weight matrix; The attention aggregate of the coordination unit is then calculated: Where D′ is the updated coordination unit embedding, which contains the information in the input O; In the distribution phase, the updated coordination unit embedding D′ performs cross-attention calculation with O again, and redistributes the aggregated information to each data slice, thereby realizing the global propagation of information; using O as the request matrix and D′ as the index and content matrix, the output after distribution is calculated; the request matrix is: Where W Q2 is the weight matrix; The coordination unit index matrix is: Where W K2 is the weight matrix; The coordination unit content matrix is: V D′ =D′W V2 ∈R k×d , where W V2 is the weight matrix; The output after calculation distribution is expressed as: O' is the information representation after redistribution. After the information propagation of the coordination unit, the information between the original data slices is further integrated.

[0013] Furthermore, in the steps S3 and S33, the coordination unit module plays a key role in the fault prediction model. It aggregates and distributes attention information by introducing a small number of coordination units, thereby effectively reducing the computational complexity. Through the coordination unit module, the model can integrate global information while reducing computing resources, ensuring that important dependencies between time and variables are fully captured. In addition, the coordination unit also helps the model maintain stability when processing large-scale data, providing more efficient and reliable support for fault prediction.

[0014] Furthermore, in the steps S3 and S34, the output O' is further processed by a layer-by-layer excitation network, and the representation of each data slice is transformed by a layer-by-layer excitation network, including two linear layers and an activation function. The specific steps are: O′→FeedForward(O′)→O″∈R N×p×d′ ; In the formula, d' is the output mapping dimension; Finally, the processed representation O″ will be mapped to the prediction space through a linear mapping layer to obtain the prediction results of each variable in the future time step; The linear mapping formula is: X out =O″W out ∈R N×S ; Where S is the number of predicted time steps, W out ∈R N×S Output linear mapping matrix, X out is the final prediction result.

[0015] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: In the present invention, a door and window fitting equipment fault prediction model under high-load operation is proposed, and the door and window fitting equipment fault prediction model under high-load operation is composed of a comprehensive preprocessing module, a fusion attention mechanism, a coordination unit module and a layer-by-layer excitation network, wherein the comprehensive preprocessing module blocks, embeds and encodes the data; the fusion attention mechanism establishes a complex dependency relationship between the time dimension and the variable dimension; the coordination unit module aggregates and distributes attention information by introducing coordination units to reduce complexity; the layer-by-layer excitation network performs a nonlinear transformation on the embedded representation of each data slice. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 The present invention is a flowchart of the steps of a method for predicting failure of door and window fitting equipment under high load operation.

[0017] Figure 2 Flowchart of the steps for developing a failure prediction model for doors and windows under high load operation.

[0018] Figure 3 Diagram of the internal structure of the fusion attention mechanism in the failure prediction model of door and window fitting equipment under high load operation.

[0019] Figure 4 This is a diagram showing the fitting of the predicted values ​​and true values ​​of the door and window fitting equipment failure prediction model under high load operation. DETAILED DESCRIPTION

[0020] 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.

[0021] See also Figure 1-Figure 4 The present invention provides a technical solution: a method for predicting failure of door and window fitting equipment under high-load operation, the method comprising the steps of collecting data related to equipment operation status and failure, normalizing the collected data, proposing a failure prediction model for door and window fitting equipment under high-load operation, training the failure prediction model for door and window fitting equipment under high-load operation, and testing the failure prediction model for door and window fitting equipment under high-load operation.

[0022] Please refer to Figure 1 As shown, a method for predicting failure of door and window fitting equipment under high-load operation in an embodiment of the present application comprises the following specific steps:

[0023] S1. Collect data related to equipment operating status and faults, and monitor vibration, temperature, current, voltage and sound parameters in real time.

[0024] Furthermore, in step S1, data related to the operating status and faults of the equipment are collected, and various sensors, such as acceleration sensors, temperature sensors, current sensors and sound sensors, are installed at key parts of the equipment to monitor vibration, temperature, current, voltage and sound parameters in real time, and a complete data acquisition system is established, including hardware and software parts, to achieve real-time monitoring, storage and preliminary processing of data; the data is transmitted to the central processing unit via wired and wireless methods, and stored in a database or big data platform, the collected data is annotated, and information such as timestamps, equipment status, load conditions, etc. is recorded, and a fault log is established to ensure the integrity and accuracy of the data; the acquisition frequency is once every 5 minutes.

[0025] S2. Normalize the collected data to eliminate the impact of different dimensions and scales.

[0026] Furthermore, in step S2, after cleaning the data collected in step S1, the cleaned data is normalized using maximum-minimum normalization so that the result value is mapped between 0 and 1. The maximum-minimum normalization formula is as follows: In the formula, x is the data before normalization, X normal is the normalized data, x min is the minimum value among the normalized attributes, x max is the maximum value among the normalized attributes.

[0027] S3. Propose a failure prediction model for door and window fitting equipment under high load operation, such as Figure 2 Shown is a step flow chart of a fault prediction model, and the specific method includes:

[0028] S31. Construct a comprehensive preprocessing module to divide, embed and encode the data.

[0029] Furthermore, in the steps S3 and S31, the time series of each variable is divided into blocks according to the window length l to obtain multiple blocks. i ∈R T It is divided into p blocks of length l, and the formula is expressed as: In the formula, is the number of blocks, each block is the jth time period; Each block is transformed into an embedded representation through linear projection. In the embedded representation, each block is first flattened and converted into a one-dimensional vector. For each block of each vector, a linear transformation W∈R is performed. l×dEmbedding it into d-dimensional space, the linear transformation formula is as follows: In the formula, is the embedding representation of the jth block, b is the bias term; after the entire data set is embedded, the embedding matrix H∈R N×p×d Block embedding for all variables; Then, position encoding is introduced to make the model have sequence order information. For the embedding of each time step, position encoding P∈R is added N×p×d , in order to retain the time order information, the added position encoding formula is: H′=H+P; Where H′ is data containing both variable and time position information.

[0030] S32. Build a fusion attention mechanism to establish complex dependencies between the time dimension and the variable dimension.

[0031] Furthermore, in the steps S3 and S32, the fusion attention mechanism establishes a complex dependency relationship between the time dimension and the variable dimension, such as Figure 3 The internal structure diagram of the fusion attention mechanism is shown. First, the input vector H′ is flattened to convert it into a one-dimensional sequence. After expansion, a new vector X′∈R is obtained. N×p×d In the fusion attention mechanism, the query matrix Q, key matrix K and value matrix V are defined to calculate the attention. In order to capture the dependencies across time and variable dimensions at the same time, the fusion attention mechanism linearly transforms the query, key and value of each block. The formula is as follows: V=X′W V ∈R N×p×d ; Where W Q , W K , W V is the trainable linear transformation matrix, N is the number of variables, p is the number of blocks for each variable, and d k is the dimension of the embedding space, which is equal to the mapping dimension d; then the output of the fusion attention mechanism is calculated for each block, and the formula is as follows: In the formula, QK T To calculate the similarity between the query and the key, Scaling is used to stabilize the gradient, and the softmax operation is used to normalize the attention weights at each position. The matrix It is the information after the interaction between the time dimension and the variable dimension.

[0032] S33. Construct a coordination unit module to aggregate and distribute attention information by introducing coordination units to reduce complexity.

[0033] Furthermore, in the steps S3 and S33, in the aggregation phase, the coordination unit performs cross attention calculation with the input O to aggregate the information in the input into the coordination unit; first, using D as the request matrix and O as the index and content matrix, the cross attention is calculated: The coordination unit request matrix is: Where k is the number of coordination units, W Q1 is the weight matrix; The input index matrix is: Where W K1 is the weight matrix; The input content matrix is: V O =OW V1 ∈R N×p×d Where W V1 is the weight matrix; The attention aggregate of the coordination unit is then calculated: Where D′ is the updated coordination unit embedding, which contains the information in the input O; In the distribution phase, the updated coordination unit embedding D′ performs cross-attention calculation with O again, and redistributes the aggregated information to each data slice, thereby realizing global propagation of information; using O as the request matrix and D′ as the index and content matrix, the output after distribution is calculated; The request matrix is: Where W Q2 is the weight matrix; The coordination unit index matrix is: Where W K2 is the weight matrix; The coordination unit content matrix is: V D′ =D′W V2 ∈R k×d , where W V2 is the weight matrix; The output after calculation distribution is expressed as: O' is the information representation after redistribution. After the information propagation of the coordination unit, the information between the original data slices is further integrated.

[0034] S34. Construct a layer-by-layer excitation network to perform nonlinear transformation on the embedded representation of each data slice.

[0035] Furthermore, in the steps S3 and S34, the output O' is further processed by a layer-by-layer excitation network, and the representation of each data slice is transformed by a layer-by-layer excitation network, including two linear layers and an activation function. The specific steps are: O′→FeedForward(O′)→O″∈R N×p×d′ ; In the formula, d' is the output mapping dimension; Finally, the processed representation O″ will be mapped to the prediction space through a linear mapping layer to obtain the prediction results of each variable in the future time step; The linear mapping formula is: X out =O″W out ∈R N×S ; Where S is the number of predicted time steps, W out ∈R N×S Output linear mapping matrix, X out is the final prediction result.

[0036] S4. Training a door and window fitting equipment failure prediction model under high-load operation, wherein the training includes hyperparameter setting and model training process.

[0037] Furthermore, in step S4, in the hyperparameter setting, the number of model training iterations Epochs is 500, the learning rate LR is 0.0001, the optimizer is Adam, and the batch size batch_size is 128; the loss function used in the model training process is the mean square error loss function.

[0038] S5. Testing the door and window fitting equipment failure prediction model under high load operation. The purpose of the test is to test whether the failure prediction model meets the requirements of the actual environment.

[0039] Furthermore, in step S5, the fault prediction model is tested using test data to determine whether the fault prediction model can be applied in an actual environment. Figure 4 As shown in the figure, the fitting of the predicted value of the prediction model with the true value is demonstrated. The data of the first 30 minutes are used to predict the failure of the door and window fitting equipment in the next 10 minutes. It can be concluded from the figure that the prediction model can accurately predict the failure in the future.

Claims

1. A method for predicting failure of door and window fitting equipment under high load operation, characterized in that: The following steps are involved: S1. Collect data related to equipment operation status and faults, and monitor vibration, temperature, current, voltage and sound parameters in real time; S2. Normalize the collected data to eliminate the dimension and scale differences between different physical quantities to ensure the consistency of model input; S3. A failure prediction model for door and window fitting equipment under high load operation is proposed. The specific method includes: S31, build a comprehensive preprocessing module to process the data in blocks, and then perform embedding and encoding steps to lay the foundation for subsequent modeling; S32. Construct a fusion attention mechanism to establish dependencies between time and variable dimensions and capture complex interactions in multivariate time series data; S33, build a coordination unit module, and reduce the computational complexity of the attention mechanism by introducing coordination units to aggregate and distribute attention information; S34, build a layer-by-layer excitation network, perform nonlinear transformation on the embedded representation of each data piece, and enhance the model's ability to express features; S4. Training a door and window fitting equipment failure prediction model under high-load operation, including hyperparameter tuning and model training steps, to ensure the accuracy of the model under high-load operation; S5. Test the failure prediction model of door and window fitting equipment under high-load operation, test the model in the actual environment, and evaluate the accuracy of the failure prediction to verify its effectiveness in actual application.

2. The method for predicting failure of door and window fitting equipment under high-load operation according to claim 1, characterized in that: In step S1, data related to the equipment operating status and faults are collected by installing various sensors, such as acceleration sensors, temperature sensors, current sensors and sound sensors, at key parts of the equipment to monitor vibration, temperature, current, voltage and sound parameters in real time. The collection frequency is once every 5 minutes.

3. The method for predicting failure of door and window fitting equipment under high-load operation according to claim 2, characterized in that: In steps S3 and S31, the time series of each variable is divided into blocks according to the window length l to obtain multiple blocks. The time series x of each variable is i ∈R T It is divided into p blocks of length l, and the formula is expressed as: In the formula, is the number of blocks, each block is the jth time period; Each block is transformed into an embedded representation through linear projection. In the embedded representation, each block is first flattened and converted into a one-dimensional vector. For each block of each vector, a linear transformation W∈R is performed. l×d Embedding it into d-dimensional space, the linear transformation formula is as follows: In the formula, is the embedding representation of the jth block, b is the bias term; after the entire data set is embedded, the embedding matrix H∈R N ×p×d Block embedding for all variables; Then, position encoding is introduced to make the model have sequence order information. For the embedding of each time step, position encoding P∈R is added N×p×d , in order to retain the time order information, the added position encoding formula is: H′=H+P; Where H′ is data containing both variable and time position information.

4. The method for predicting failure of door and window fitting equipment under high-load operation according to claim 3, characterized in that: In steps S3 and S32, the fusion attention mechanism establishes a complex dependency relationship between the time dimension and the variable dimension. First, the input vector H′ is flattened to convert it into a one-dimensional sequence. After expansion, a new vector X′∈R is obtained. N×p×d In the fusion attention mechanism, the query matrix Q, key matrix K and value matrix V are defined to calculate the attention. In order to capture the dependencies across time and variable dimensions at the same time, the fusion attention mechanism linearly transforms the query, key and value of each block. The formula is as follows: V=X′W V ∈R N×p×d ; Where WQ 、 W K , W V is the trainable linear transformation matrix, N is the number of variables, p is the number of blocks for each variable, and d k is the dimension of the embedding space, which is equal to the mapping dimension d; then the output of the fusion attention mechanism is calculated for each block, and the formula is as follows: In the formula, QK T To calculate the similarity between the query and the key, Scaling is used to stabilize the gradient, and the softmax operation is used to normalize the attention weights at each position. The matrix It is the information after the interaction between the time dimension and the variable dimension.

5. The method for predicting failure of door and window cutting equipment under high-load operation according to claim 4, characterized in that: In steps S3 and S33, in the aggregation phase, the coordination unit performs cross attention calculation with the input O to aggregate the information in the input into the coordination unit; first, using D as the request matrix and O as the index and content matrix, the cross attention is calculated: the coordination unit request matrix is: Where k is the number of coordination units, W Q1 is the weight matrix; The input index matrix is: Where W K1 is the weight matrix; The input content matrix is: V O =OW V1 ∈R N×p×d Where W V1 is the weight matrix; The attention aggregate of the coordination unit is then calculated: Where D′ is the updated coordination unit embedding, which contains the information in the input O; In the distribution phase, the updated coordination unit embedding D′ performs cross-attention calculation with O again, and redistributes the aggregated information to each data slice, thereby realizing the global propagation of information; using O as the request matrix and D′ as the index and content matrix, the output after distribution is calculated; the request matrix is: Where W Q2 is the weight matrix; The coordination unit index matrix is: Where W K2 is the weight matrix; The coordination unit content matrix is: V D′ =D′W V2 ∈R k×d , where W V2 is the weight matrix; The output after calculation distribution is expressed as: O′ is the information representation after redistribution. After the information propagation of the coordination unit, the information between the original data slices is further integrated.

6. The method for predicting failure of door and window cutting equipment under high-load operation according to claim 5, characterized in that: In steps S3 and S34, the output O′ is further processed through a layer-by-layer excitation network. The representation of each data slice is transformed through a layer-by-layer excitation network, including two linear layers and an activation function. The specific steps are: O′→FeedForward(O′)→O″∈R N×p×d′ ; Where d′ is the output mapping dimension; Finally, the processed representation O″ will be mapped to the prediction space through a linear mapping layer to obtain the prediction results of each variable in the future time step; the linear mapping formula is: X out =(O″W out ∈R N×S ; Where S is the number of time steps for prediction, W out ∈R N×S Output linear mapping matrix, X out is the final prediction result.

Images