Industrial equipment service life prediction method based on large model knowledge enhancement

Abstract

The invention relates to the technical field of equipment predictive maintenance, and particularly provides an industrial equipment service life prediction method based on large model knowledge enhancement. The method comprises the following steps: collecting operation monitoring data of equipment, and processing the operation monitoring data into log text data and time sequence data according to types; performing multi-modal data remodeling on the log text data and the time sequence data to obtain a sequence block; according to the method, the sequence blocks are embedded into the dimension space to predict the service life of the equipment, and the accuracy and generalization ability of prediction are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment predictive maintenance, and in particular to a method for predicting the service life of industrial equipment based on large model knowledge enhancement. Background Art

[0002] In the industrial field, equipment life prediction is crucial for maintenance planning and cost control. Traditional prediction methods rely on a large amount of historical data and complex physical models, but in practical applications, it is often difficult to obtain sufficient data, especially for new equipment or equipment in special environments. In addition, traditional methods are insufficient in prediction accuracy and generalization ability when dealing with small sample data. Summary of the invention

[0003] In view of this, the present invention provides an industrial equipment service life prediction method based on large model knowledge enhancement to improve the accuracy and generalization ability of prediction.

[0004] In a first aspect, the present invention provides a method for predicting the service life of industrial equipment based on large model knowledge enhancement, the method comprising:

[0005] Step 1: Collect the operation monitoring data of the equipment, and process the operation monitoring data into log text data and time series data according to the type;

[0006] Step 2: Perform multimodal data reshaping on the log text data and time series data, train the reshaping module, and obtain sequence blocks;

[0007] Step 3: Embed the sequence blocks into the dimensional space and embed the head-enhanced text prompts into the pre-trained large language model to predict the equipment life.

[0008] Optionally, step 1 includes:

[0009] For log text data: remove useless symbols, stop words, punctuation marks, perform stem extraction or word form restoration; split the log text data into tokens of words, phrases or sub-word units in time units, and the expression is: T = Tokenize (S); where S represents the original text, T represents the token sequence after word segmentation, and Tokenize () represents the word segmentation operation;

[0010] For time series data: Divide into single channels according to variables, and normalize each input channel so that its mean is 0 and its standard deviation is 1.

[0011] Optionally, the step 2 includes:

[0012] For log text data: convert the segmented log text data into an embedding vector, that is, text embedding; divide the segmented token sequence into blocks of fixed length. Assume that the total sequence is S, the length of the input sequence is L, and the input block p text The total length is Convert each text block into an embedding vector, which is done through a pre-trained embedding layer, and its expression is: Where E embedding is the embedding function;

[0013] For time series data: normalize the time series data P time Split into multiple continuous, overlapping or non-overlapping sequence blocks; assuming the total sequence is S, the length of the input sequence is L, the output prediction sequence length is F, and the input sequence block p t The total number is Each normalized sequence block is embedded into dimension d m In the space of , each sequence block represents a local segment of the time series data.

[0014] Optionally, the step 3 includes:

[0015] The sequence blocks are embedded into a high-dimensional space, using a linear layer as the sequence block embedder; the embedding is expressed as: Individual variables are normalized and scored, and sequence blocks are embedded into a dimensional space via a learnable embedding layer.

[0016] Optionally, in order to align the time series data with natural language, create a set of time series-text vocabulary, extract the text vocabulary from the pre-trained word embeddings to represent the characteristics of the time series data;

[0017] Assume that the vocabulary of the pre-trained large language model is V, which contains N words, each of which has a corresponding embedding vector; select a subset V′ from the vocabulary, which contains M words; in order to map the time series data to the pre-trained text embedding space, use a linear layer to detect the text vocabulary, and the expression of the process is:

[0018] E′ i =E h(x) +α·(xE h(x) );

[0019] Where h(x) is a hash function used to map the feature vector of time series data to a word in the vocabulary V; E h(x) is the embedding vector of the word corresponding to h(x) in the vocabulary V; α is a scalar used to adjust the step size of linear detection; E′ i is the embedding vector corresponding to word i in subset V′;

[0020] Linear detection is performed on multiple feature vectors to obtain the embedding vector that describes the time series data. The expression of the process is:

[0021]

[0022] The above formula considers the mapping between the feature vector of time series data and the pre-trained text embedding space, and adjusts the embedding vector through linear detection so that it can describe the characteristics of time series data;

[0023] Use a multi-head cross-attention mechanism to aggregate time series vocabulary embedding and text prototype embedding, which is expressed as:

[0024]

[0025] Where Q(i) k is the query matrix, used to represent the time series embedding; K(i) k and V(i) k are key and value matrices, respectively, used to represent the text vocabulary embedding; and is the weight matrix for each attention head; d k is the dimension of each attention head; b is the bias term;

[0026] Introduce nonlinear activation functions to increase the nonlinear capabilities of the model;

[0027] Aggregate the output of each attention head and linearly project it to the hidden layer dimension of the large language model, which is expressed as:

[0028] Z(i)=CONCAT(Z(i)1,Z(i)2,...,Z(i) K );

[0029] The final output is sent to a large language model for further processing;

[0030] The embedding vectors of log text data and time series data are concatenated along the feature dimension to obtain the embedding matrix Z, which is expressed as: Z = Concat(E text , E time );

[0031] The above operations convert the sequence block embeddings of time series into text representations to activate the time series understanding and reasoning capabilities of large language models.

[0032] Optionally, the head augments the text prompts to enhance the reasoning ability of the large language model by adding additional context and task instructions to the input time sequence, guiding the large language model to transform the input sequence blocks.

[0033] Optionally, language cues relevant to device lifespan prediction are defined and fed into a pre-trained large language model as head-augmented text cues along with the aligned cross-modal data.

[0034] In order to enhance the reasoning ability of large language models and improve their processing efficiency for time series data, dynamic context generation is used to optimize the head-enhanced text prompts, and context information is dynamically generated according to the characteristics of time series data. The trend characteristics and periodic characteristics of time series data are analyzed through deep learning algorithms, and text descriptions corresponding to the trend characteristics and periodic characteristics are generated and input into the model as context information.

[0035] Optionally, a large language model is pre-trained and used to predict service life. The parameters of the pre-trained large language model are frozen, that is, they are not updated during the prediction process. The large language model receives the embedding matrix Z of the spliced ​​log text data and time series data in the multimodal data reshaping and the life prediction language prompt input; the pre-trained large language model is used to process the input cross-modal data and language prompts to generate service life prediction data.

[0036] Optionally, prediction projection, flattens and linearly projects the output representation processed by the large language model to generate a final prediction result, i.e., the useful life in units of time.

[0037] Optionally, the loss function and labels used for training include:

[0038] (1) Uncertain penalty and cost maintenance loss function:

[0039] Let y i The known service life of the i-th device is the true value, is the service life predicted by the model, that is, the predicted value, σ i is the uncertainty of prediction, that is, the standard deviation, and C is the maintenance cost coefficient of the equipment;

[0040] Define a new loss function L, which is expressed as:

[0041]

[0042] Among them, w1, w2, and w3 are weight coefficients used to balance the influence of each item; is the squared error term, penalizing the difference between the predicted value and the true value; A penalty term for prediction uncertainty, used to encourage the model to output certain predictions; It is the advanced maintenance cost item, which is used to penalize the situation where the service life predicted by the model exceeds the actual service life. It is multiplied by the maintenance cost coefficient C to reflect the actual economic loss.

[0043] (2) Tags:

[0044] Collect data from the equipment's historical maintenance records, including the equipment's installation date, failure date, and replacement date. The calculation method for its service life label is: service life = end point date - start point date. For retired equipment, record the total time from commissioning to retirement as the actual service life of the equipment.

[0045] Few-sample data enhancement uses the collected data to generate new data samples by adding noise and interpolating time series data, which includes:

[0046] a. By introducing the time-dependent noise standard deviation, the noise changes of the equipment at different use stages are simulated, so that the model can adapt to the dynamic changes of the equipment status;

[0047] Let x t is the original time series data, ∈ is the injected noise, where σ t is the time-dependent noise standard deviation:

[0048] x′ t =x t +∈

[0049] σ t =σ0·e -λt ;

[0050] Where σ0 is the initial noise level, λ is the decay rate, and t is the time index;

[0051] b. By introducing time-dependent interpolation coefficients, we can simulate the changing trend of the equipment in different time periods and enhance the sensitivity of the model to changes in equipment status:

[0052] x′ t =α t x t +(1-α t )x t+1

[0053]

[0054] α t is the time-dependent interpolation coefficient, γ is the adjustment coefficient, and T is the midpoint time index;

[0055] c. Introducing a time-dependent scaling factor to dynamically adjust the influence of noise in fault mode synthesis, so that the model can better capture fault characteristics in different time periods;

[0056] Assume f is the fault feature vector, g is the function of synthetic fault mode, z is the noise vector, βt is the time-dependent scaling factor, δ is the adjustment rate, and t is the time index, then:

[0057] f′=g(f,z)=f+β t ·z

[0058] β t =β0tanh(δt).

[0059] In the technical solution provided by the present invention, the method includes collecting operation monitoring data of the equipment, processing the operation monitoring data into log text data and time series data by type; performing multimodal data reshaping on the log text data and the time series data to obtain sequence blocks; embedding the sequence blocks into a dimensional space to predict the service life of the equipment. The method improves the accuracy and generalization ability of the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0061] Figure 1 A flowchart of a method for predicting the service life of industrial equipment provided by an embodiment of the present invention;

[0062] Figure 2 A flowchart of time series data reshaping provided by an embodiment of the present invention;

[0063] Figure 3 A schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution 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 part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0065] It should be clear that the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0066] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. The singular forms "a", "said" and "the" used in the embodiments of the present invention are also intended to include plural forms, unless the context clearly indicates other meanings.

[0067] It should be understood that the term "and / or" used in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0068] The word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to determining" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)", depending on the context.

[0069] Large language models such as ChatGPT learn the basic structure and patterns of language by pre-training on a large amount of text data, and then fine-tune through supervised learning and reinforcement learning to adapt to specific application scenarios and tasks. It can achieve a variety of natural language processing tasks including text generation, automatic question answering, language translation, content summarization, sentiment analysis, text classification, dialogue system construction, as well as cross-domain applications such as auxiliary programming, education, and writing. Large language models can learn rich language patterns and structures through pre-training on a large amount of text data, which enables them to better understand and analyze changes in device status and performance when processing device monitoring data, thereby improving the accuracy of predictions.

[0070] By using data enhancement techniques in natural language processing, such as synonym replacement and back translation, more training samples can be generated to improve model performance and the generalization ability of the model on monitoring data. Through natural language processing technology, key information can be extracted from a large amount of text data for training and optimization of prediction models, thereby improving the accuracy of equipment service life prediction.

[0071] Figure 1 A flowchart of a method for predicting the service life of industrial equipment provided by an embodiment of the present invention is shown in FIG. Figure 1 As shown, the method includes:

[0072] Step 1: Collect the operation monitoring data of the equipment, and process the operation monitoring data into log text data and time series data by type.

[0073] In the embodiment of the present invention, step 1 includes:

[0074] For log text data: remove useless symbols, stop words, punctuation marks, perform stem extraction or word form restoration; split the log text data into tokens of words, phrases or sub-word units in time units, and the expression is: T = Tokenize (S); where S represents the original text, T represents the token sequence after word segmentation, and Tokenize () represents the word segmentation operation;

[0075] For time series data: Divide into single channels according to variables, and normalize each input channel so that its mean is 0 and its standard deviation is 1.

[0076] Step 2: Perform multimodal data reshaping on the log text data and time series data, train the reshaping module, and obtain sequence blocks.

[0077] In the embodiment of the present invention, step 2 includes:

[0078] For log text data: convert the segmented log text data into an embedding vector, that is, text embedding; divide the segmented token sequence into blocks of fixed length. Assume that the total sequence is S, the length of the input sequence is L, and the input block p text The total length is Convert each text block into an embedding vector, which is done through a pre-trained embedding layer, and its expression is: Where E embedding is the embedding function;

[0079] In the embodiment of the present invention, the segmented log text data is converted into embedding vectors, and these embeddings can be pre-trained word embeddings or embeddings obtained through model learning.

[0080] For time series data: normalize the time series data P time Split into multiple continuous, overlapping or non-overlapping sequence blocks; assuming the total sequence is S, the length of the input sequence is L, the output prediction sequence length is F, and the input sequence block p t The total number is Each normalized sequence block is embedded into dimension d m In the space of , each sequence block represents a local segment of the time series data.

[0081] Step 3: Embed the sequence blocks into the dimensional space and embed the head-enhanced text prompts into the pre-trained large language model to predict the equipment life.

[0082] In the embodiment of the present invention, Figure 2 As shown, step 3 includes:

[0083] The sequence blocks are embedded into a high-dimensional space, using a linear layer as the sequence block embedder; the embedding is expressed as: Individual variables are normalized and scored, and sequence blocks are embedded into a dimensional space via a learnable embedding layer.

[0084] In an embodiment of the present invention, in order to align time series data with natural language, a set of time series-text vocabulary libraries is created, and a text vocabulary library is extracted from pre-trained word embeddings to represent the characteristics of time series data; the text in the vocabulary library is a vocabulary that describes the characteristics of the time series, such as rising, falling, stable, fluctuating, rapid, etc.

[0085] Assume that the vocabulary of the pre-trained large language model is V, which contains N words, each of which has a corresponding embedding vector; select a subset V′ from the vocabulary, which contains M words; these words are selected because they are semantically related to the characteristics of the device detection data (such as trends, periodicity, etc.) and have the most information when describing device dynamics. In order to map time series data to the pre-trained text embedding space, a linear layer is used to detect the text vocabulary, and the expression of the process is:

[0086] E′ i =E h(x) +α·(xE h(x) );

[0087] Where h(x) is a hash function used to map the feature vector of time series data to a word in the vocabulary V; E h(x) is the embedding vector of the word corresponding to h(x) in the vocabulary V; α is a scalar used to adjust the step size of linear detection; E′ i is the embedding vector corresponding to word i in subset V′;

[0088] Linear detection is performed on multiple feature vectors to obtain the embedding vector that describes the time series data. The expression of the process is:

[0089]

[0090] The above formula considers the mapping between the feature vector of time series data and the pre-trained text embedding space, and adjusts the embedding vector through linear detection so that it can describe the characteristics of time series data;

[0091] Use a multi-head cross-attention mechanism to aggregate time series vocabulary embedding and text prototype embedding, which is expressed as:

[0092]

[0093] Where Q(i) k is the query matrix, used to represent the time series embedding; K(i)k and V(i) k are key and value matrices, respectively, used to represent the text vocabulary embedding; and is the weight matrix for each attention head; d k is the dimension of each attention head; b is the bias term;

[0094] Introduce nonlinear activation functions to increase the nonlinear capabilities of the model;

[0095] Aggregate the output of each attention head and linearly project it to the hidden layer dimension of the large language model, which is expressed as:

[0096] Z(z)=CONCAT(Z(i)1,Z(i)2,...,Z(i) K );

[0097] The final output is sent to a large language model for further processing;

[0098] The embedding vectors of log text data and time series data are concatenated along the feature dimension to obtain the embedding matrix Z, which is expressed as: Z = Concat(E text , E time );

[0099] The above operations convert the sequence block embeddings of time series into text representations to activate the time series understanding and reasoning capabilities of large language models.

[0100] In the embodiment of the present invention, the head-enhanced text prompts enhance the reasoning ability of the large language model by adding additional context and task instructions to the input time sequence, and guide the large language model to transform the input sequence blocks. In the embodiment of the present invention, language prompts related to equipment service life prediction are defined and input into the pre-trained large language model as head-enhanced text prompts together with the aligned cross-modal data.

[0101] In order to enhance the reasoning ability of the large language model and improve its processing efficiency of time series data, dynamic context generation is used to optimize the head enhanced text prompts, and context information is dynamically generated according to the characteristics of the time series data, which further improves the reasoning ability and processing efficiency of the large language model; the trend characteristics and periodic characteristics of the time series data are analyzed through deep learning algorithms, and text descriptions corresponding to the trend characteristics and periodic characteristics are generated, and they are input into the model as context information.

[0102] By combining natural language processing technology and deep learning technology, the present invention can extract key information from unstructured natural language text and combine it with equipment monitoring data. Test results in real scenarios show that the accuracy of equipment service life prediction is improved by 28%.

[0103] In an embodiment of the present invention, a large language model is pre-trained, and the large language model is used to predict the service life. The parameters of the pre-trained large language model are frozen, that is, they are not updated during the prediction process; the large language model receives the embedding matrix Z of the spliced ​​log text data and time series data in the multimodal data reshaping and the life prediction language prompt input; the pre-trained large language model is used to process the input cross-modal data and language prompts to generate service life prediction data.

[0104] In the embodiment of the present invention, the prediction projection flattens and linearly projects the output representation processed by the large language model to generate a final prediction result, that is, the service life in units of time.

[0105] In the embodiment of the present invention, the loss function and labels used for training include:

[0106] (1) Uncertain penalty and cost maintenance loss function:

[0107] Let y i The known service life of the i-th device is the true value, is the service life predicted by the model, that is, the predicted value, σ i is the uncertainty of prediction, that is, the standard deviation, and C is the maintenance cost coefficient of the equipment;

[0108] Define a new loss function L, which is expressed as:

[0109]

[0110] Among them, w1, w2, and w3 are weight coefficients used to balance the influence of each item; is the squared error term, penalizing the difference between the predicted value and the true value; A penalty term for prediction uncertainty, used to encourage the model to output certain predictions; It is the advanced maintenance cost item, which is used to penalize the situation where the service life predicted by the model exceeds the actual service life. It is multiplied by the maintenance cost coefficient C to reflect the actual economic loss.

[0111] The weight coefficient needs to be adjusted according to the specific application scenario and data characteristics. The optimal weight combination can be found through methods such as cross-validation. The squared error term ensures that the model pays enough attention to the accuracy of the prediction. The added penalty term for prediction uncertainty encourages the model to take into account its own uncertainty when predicting, thereby improving the reliability of the model. By introducing the maintenance cost coefficient, the model takes into account the economic losses caused by premature replacement of equipment when predicting.

[0112] (2) Tags:

[0113] Collect data from the equipment's historical maintenance records, including the equipment's installation date, failure date, and replacement date. The calculation method for its service life label is: service life = end point date - start point date. For retired equipment, record the total time from commissioning to retirement as the actual service life of the equipment.

[0114] In the case of limited equipment data samples, the knowledge enhancement capability of the pre-trained model is used to improve the generalization ability and prediction accuracy of the model. Combined with the semantic understanding ability of the pre-trained model, natural language descriptions related to equipment status and maintenance history are generated as part of data enhancement. The enhanced data is used for model training to improve the model's ability to recognize different equipment states and failure modes.

[0115] Few-sample data enhancement uses the collected data to generate new data samples by adding noise and interpolating time series data, which includes:

[0116] a. By introducing the time-dependent noise standard deviation, the noise changes of the equipment at different use stages are simulated, so that the model can adapt to the dynamic changes of the equipment status;

[0117] Let x t is the original time series data, ∈ is the injected noise, where σ t is the time-dependent noise standard deviation:

[0118] x′ t =x t +∈

[0119] σ t =σ0·e -λt ;

[0120] Where σ0 is the initial noise level, λ is the decay rate, and t is the time index;

[0121] b. By introducing time-dependent interpolation coefficients, we can simulate the changing trend of the equipment in different time periods and enhance the sensitivity of the model to changes in equipment status:

[0122] x′ t =αt x t +(1-α t )x t+1

[0123]

[0124] α t is the time-dependent interpolation coefficient, γ is the adjustment coefficient, and T is the midpoint time index;

[0125] c. Introducing a time-dependent scaling factor to dynamically adjust the influence of noise in fault mode synthesis, so that the model can better capture fault characteristics in different time periods;

[0126] Assume f is the fault feature vector, g is the function of synthetic fault mode, z is the noise vector, β t is the time-dependent scaling factor, δ is the adjustment rate, and t is the time index, then:

[0127] f′=g(f,z)=f+β t ·z

[0128] β t =β0·tanh(δt).

[0129] Cross-modal data alignment technology enables continuous equipment monitoring data to be aligned with natural language descriptions, enhancing the model's ability to understand equipment status and achieving seamless integration with large language models. By generating natural language descriptions related to equipment status and maintenance history, the present invention uses the knowledge enhancement capabilities of pre-trained models to improve the model's ability to recognize different equipment states and failure modes when equipment data samples are limited. By using existing equipment data, new data samples are generated by adding noise and interpolating time series data, which increases the number of training samples and improves the model's prediction effect.

[0130] Compared with the prior art, the present invention has the following beneficial effects:

[0131] 1. A cross-modal alignment method for reshaping monitoring data is used. This method can convert continuous equipment monitoring data (such as sensor readings and time series data) into natural language descriptions and align them with natural language to effectively utilize the text generation capabilities of the large language model to generate cross-modal industrial equipment operation prediction data, achieving seamless integration with the large language model. This alignment technology not only improves the readability of the data, but also enhances the model's ability to understand the equipment status.

[0132] 2. A large-model equipment life prediction method based on pre-trained knowledge enhancement is adopted, and natural language prompt technology is introduced to enhance the reasoning ability of the large model in the equipment life prediction task by designing specific language prompts. The service life prediction data is generated by inputting the aligned cross-modal data and specific service life prediction language prompts into the pre-trained large language model. This technology enables the model to more accurately extract key information from unstructured text and time series data, and combine it with equipment monitoring data to improve the accuracy of prediction.

[0133] 3. A few-sample equipment service life prediction method is adopted. In view of the problem of limited equipment data samples, the present invention proposes a few-sample learning optimization technology, which enhances the knowledge capability of the pre-training model and improves the generalization ability and prediction accuracy of the model under limited samples.

[0134] In the technical solution provided by the present invention, the method includes collecting operation monitoring data of the equipment, processing the operation monitoring data into log text data and time series data by type; performing multimodal data reshaping on the log text data and the time series data to obtain sequence blocks; embedding the sequence blocks into a dimensional space to predict the service life of the equipment. The method improves the accuracy and generalization ability of the prediction.

[0135] Each step of the embodiment of the present invention may be performed by an electronic device, which includes but is not limited to a mobile phone, a tablet computer, a portable PC, a desktop computer, etc.

[0136] An embodiment of the present invention provides a computer-readable storage medium, which includes a stored program, wherein when the program is running, the electronic device where the computer-readable storage medium is located is controlled to execute the above-mentioned embodiment of the industrial equipment service life prediction method based on large model knowledge enhancement.

[0137] Figure 3 A schematic diagram of an electronic device provided by an embodiment of the present invention, such as Figure 3 As shown, the electronic device 21 includes: a processor 211, a memory 212, and a computer program 213 stored in the memory 212 and executable on the processor 211. When the computer program 213 is executed by the processor 211, the industrial equipment service life prediction method based on large model knowledge enhancement in the embodiment is implemented. To avoid repetition, they are not described one by one here.

[0138] The electronic device 21 includes, but is not limited to, a processor 211 and a memory 212. Those skilled in the art will appreciate that Figure 3It is only an example of the electronic device 21 and does not constitute a limitation of the electronic device 21. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the electronic device may also include input and output devices, network access devices, buses, etc.

[0139] The processor 211 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0140] The memory 212 may be an internal storage unit of the electronic device 21, such as a hard disk or memory of the electronic device 21. The memory 212 may also be an external storage device of the electronic device 21, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the electronic device 21. Further, the memory 212 may also include both an internal storage unit of the electronic device 21 and an external storage device. The memory 212 is used to store computer programs and other programs and data required by network devices. The memory 212 may also be used to temporarily store data that has been output or is to be output.

[0141] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0142] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for predicting the service life of industrial equipment based on large model knowledge enhancement, characterized in that: The method comprises: Step 1: Collect the operation monitoring data of the equipment, and process the operation monitoring data into log text data and time series data according to the type; Step 2: Perform multimodal data reshaping on the log text data and time series data, train the reshaping module, and obtain sequence blocks; Step 3: Embed the sequence blocks into the dimensional space and embed the head-enhanced text prompts into the pre-trained large language model to predict the equipment life.

2. The method according to claim 1, characterized in that The step 1 comprises: For log text data: remove useless symbols, stop words, punctuation marks, perform stem extraction or word form restoration; split the log text data into tokens of words, phrases or sub-word units in time units, and the expression is: T = Tokenize (S); where S represents the original text, T represents the token sequence after word segmentation, and Tokenize () represents the word segmentation operation; For time series data: Divide into single channels according to variables, and normalize each input channel so that its mean is 0 and its standard deviation is 1.

3. The method according to claim 1, characterized in that The step 2 comprises: For log text data: convert the segmented log text data into an embedding vector, that is, text embedding; divide the segmented token sequence into blocks of fixed length. Assume that the total sequence is S, the length of the input sequence is L, and the input block p text The total length is Convert each text block into an embedding vector, which is done through a pre-trained embedding layer, and its expression is: Where E embedding is the embedding function; For time series data: normalize the time series data P time Split into multiple continuous, overlapping or non-overlapping sequence blocks; assuming the total sequence is S, the length of the input sequence is L, the output prediction sequence length is F, and the input sequence block p t The total number is Each normalized sequence block is embedded into dimension d m In the space of , each sequence block represents a local segment of the time series data.

4. The method according to claim 1, characterized in that: The step 3 comprises: The sequence blocks are embedded into a high-dimensional space, using a linear layer as the sequence block embedder; the embedding is expressed as: Individual variables are normalized and scored, and sequence blocks are embedded into a dimensional space via a learnable embedding layer.

5. The method according to claim 4, characterized in that In order to align time series data with natural language, a set of time series-text vocabulary is created, and the text vocabulary is extracted from the pre-trained word embedding to represent the characteristics of time series data; Assume that the vocabulary of the pre-trained large language model is V, which contains N words, each of which has a corresponding embedding vector; select a subset V from the vocabulary, which contains M words; in order to map the time series data to the pre-trained text embedding space, use a linear layer to detect the text vocabulary, and the expression of the process is: AND' i =And h(x) +α·(xE h(x) ); Where h(x) is a hash function used to map the feature vector of time series data to a word in the vocabulary V; E h(x) is the embedding vector of the word corresponding to h(x) in the vocabulary V; α is a scalar used to adjust the step size of linear detection; E′ i is the embedding vector corresponding to word i in subset V′; Linear detection is performed on multiple feature vectors to obtain the embedding vector that describes the time series data. The expression of the process is: The above formula considers the mapping between the feature vector of time series data and the pre-trained text embedding space, and adjusts the embedding vector through linear detection so that it can describe the characteristics of time series data; Use a multi-head cross-attention mechanism to aggregate time series vocabulary embedding and text prototype embedding, which is expressed as: Where Q(i) k is the query matrix, used to represent the time series embedding; K(i) k and V(i) k are key and value matrices, respectively, used to represent the text vocabulary embedding; and is the weight matrix for each attention head; d k is the dimension of each attention head; b is the bias term; Introduce nonlinear activation functions to increase the nonlinear capabilities of the model; Aggregate the output of each attention head and linearly project it to the hidden layer dimension of the large language model, which is expressed as: Z(i)=CONCAT(Z(i)1,Z(i)2,...,Z(i) K ); The final output is sent to a large language model for further processing; The embedding vectors of log text data and time series data are concatenated along the feature dimension to obtain the embedding matrix Z, which is expressed as: Z = Concat(E text , E time ); The above operations convert the sequence block embeddings of time series into text representations to activate the time series understanding and reasoning capabilities of large language models.

6. The method according to claim 5, characterized in that Head-enhanced text prompts enhance the reasoning ability of the large language model by adding additional context and task instructions to the input time sequence, guiding the large language model to transform the input sequence blocks.

7. The method according to claim 6, characterized in that Linguistic cues relevant to device lifespan prediction are defined and fed into a pre-trained large language model as head-augmented text cues along with aligned cross-modal data. In order to enhance the reasoning ability of large language models and improve their processing efficiency for time series data, dynamic context generation is used to optimize the head-enhanced text prompts, and context information is dynamically generated according to the characteristics of time series data. The trend characteristics and periodic characteristics of time series data are analyzed through deep learning algorithms, and text descriptions corresponding to the trend characteristics and periodic characteristics are generated and input into the model as context information.

8. The method according to claim 5, characterized in that Pre-train a large language model and use it to predict service life. Freeze the parameters of the pre-trained large language model, that is, do not update it during the prediction process. The large language model receives the embedding matrix Z of the spliced ​​log text data and time series data in the multimodal data reshaping and the life prediction language prompt input. Use the pre-trained large language model to process the input cross-modal data and language prompts to generate service life prediction data.

9. The method according to claim 8, characterized in that Prediction projection, flattens and linearly projects the output representation processed by the large language model to generate the final prediction result, which is the service life in time units.

10. The method according to claim 8, characterized in that The loss functions and labels used for training include: (1) Uncertain penalty and cost maintenance loss function: Let y i The known service life of the i-th device is the true value, is the service life predicted by the model, that is, the predicted value, σ i is the uncertainty of prediction, that is, the standard deviation, and C is the maintenance cost coefficient of the equipment; Define a new loss function L, which is expressed as: Among them, w1, w2, and w3 are weight coefficients used to balance the influence of each item; is the squared error term, penalizing the difference between the predicted value and the true value; A penalty term for prediction uncertainty, used to encourage the model to output certain predictions; It is the advanced maintenance cost item, which is used to penalize the situation where the service life predicted by the model exceeds the actual service life. It is multiplied by the maintenance cost coefficient C to reflect the actual economic loss. (2) Tags: Collect data from the equipment's historical maintenance records, including the equipment's installation date, failure date, and replacement date. The calculation method for its service life label is: service life = end point date - start point date. For retired equipment, record the total time from commissioning to retirement as the actual service life of the equipment. Few-sample data enhancement uses the collected data to generate new data samples by adding noise and interpolating time series data, which includes: a. By introducing the time-dependent noise standard deviation, the noise changes of the equipment at different use stages are simulated, so that the model can adapt to the dynamic changes of the equipment status; Let x t is the original time series data, ∈ is the injected noise, where σ t is the time-dependent noise standard deviation: x′ t =x t +∈ s t =σ0·e -λt ; Where σ0 is the initial noise level, λ is the decay rate, and t is the time index; b. By introducing time-dependent interpolation coefficients, we can simulate the changing trend of the equipment in different time periods and enhance the sensitivity of the model to changes in equipment status: x′ t =a t x t +(1-a t )x t+1 α t is the time-dependent interpolation coefficient, γ is the adjustment coefficient, and T is the midpoint time index; c. Introducing a time-dependent scaling factor to dynamically adjust the influence of noise in fault mode synthesis, so that the model can better capture fault characteristics in different time periods; Assume f is the fault feature vector, g is the function of synthetic fault mode, z is the noise vector, β t is the time-dependent scaling factor, δ is the adjustment rate, and t is the time index, then: f′=g(f,z)=f+β t ·z β t =β0·tanh(δt)。

Images