A device residual life prediction method, system, electronic device and storage medium

Abstract

The present application relates to the technical field of equipment health management and residual life prediction, and specifically discloses a kind of equipment residual life prediction method, system, electronic equipment and storage medium.The present application is by to the sliding division of multivariate monitoring time series and constructs structured degradation text, utilizes cross-modal contrast pre-training to map time series representation and text representation to shared semantic space, and obtains prototype semantic vector in the interval sampling ordered prototype value of residual life label, generates dynamic soft prompt in combination with fuzzy membership distribution, inputs after context organization of time series embedding and text evidence into the large language model of fine-tuning adapted, solve the problem that time series degradation feature and health state semantics cannot be uniformly modeled, and the problems of insufficient long-term prediction stability and limited cross-domain generalization in few-sample, cross-condition and cross-device scenarios are solved, the deep alignment of time series features and text semantics is realized, and the residual life prediction accuracy and numerical output stability of complex equipment under changing conditions are improved.

Description

Technical Field

[0001] This invention relates to the field of equipment health management and remaining life prediction technology, and in particular to a method, system, electronic device, and storage medium for predicting the remaining life of equipment. Background Technology

[0002] Predicting the remaining life of equipment is crucial for ensuring the safe operation of critical equipment and reducing maintenance costs. Existing methods, particularly mechanism-based modeling, typically rely on object-specific prior knowledge and precise degradation models, making them ill-suited for complex equipment and changing operating conditions. While data-driven methods possess strong nonlinear modeling capabilities, they largely depend on sufficient labeled samples, and their predictive performance tends to degrade significantly when there are differences in the distribution between the source and target domains.

[0003] Transfer learning and domain adaptation methods can alleviate the problems of data scarcity and distribution bias to some extent, but they are still insufficient for unified modeling of the relationship between temporal degradation features, health state semantics, and continuous lifetime labels. Existing large language model methods mostly use numerical signals and time-frequency features as inputs, and text semantics do not directly participate in the prediction process, which still has limitations in long-term prediction, cross-domain generalization, and numerical prediction stability.

[0004] Therefore, how to achieve joint modeling of temporal degradation features and health status semantics under conditions of few samples, cross-operating conditions, and cross-equipment, and improve the stability of long-term prediction, cross-domain generalization, and numerical prediction, remains an urgent technical problem to be solved. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method, system, electronic device, and storage medium for predicting the remaining lifespan of a device.

[0006] In a first aspect, the present invention provides a method for predicting the remaining life of equipment, the technical solution of which is as follows: The multivariate monitoring time series data of the device to be predicted is subjected to sliding segmentation to obtain multiple sample windows, and the structured degradation text corresponding to each sample window is constructed. Each sample window is input into the temporal encoder to obtain a temporal representation. Each structured degenerate text is input into the pre-trained text encoder and mapped to the shared semantic space through a projection layer to obtain a text representation. A contrastive loss is constructed based on the matched and unmatched temporal-text pairs. The temporal encoder is then pre-trained across modal contrastive modes to obtain the pre-trained temporal encoder. Finally, the pre-trained text encoder is frozen. Multiple ordered prototype values ​​are sampled in the remaining lifetime label interval and converted into text anchors. Each text anchor is encoded into a prototype semantic vector using the frozen pre-trained text encoder. For any sample window, the sample window is input into the pre-trained temporal encoder to obtain the corresponding target temporal representation. The matching score between the target temporal representation and each prototype semantic vector is calculated and obtained by the softmax function. The fuzzy membership distribution is input into the prompt learner to obtain dynamic soft prompts. The target temporal representation is projected into the temporal embedding of a large language model. The dynamic soft prompts, the temporal embedding, textual evidence, and task instructions are organized into a conditional input context. The conditions are input into the context of a large language model that has been efficiently fine-tuned with parameters to generate a structured response. The numerical fields in the structured response are then parsed to obtain the remaining lifetime prediction result of the device to be predicted.

[0007] The beneficial effects of the equipment remaining life prediction method of the present invention are as follows: The method of this invention performs sliding segmentation on multivariate monitoring time series and constructs structured degradation text. It uses cross-modal contrastive pre-training to map the temporal representation and text representation to a shared semantic space, and samples ordered prototype values ​​in the remaining lifetime label interval to obtain prototype semantic vectors. Combined with fuzzy membership distribution, it generates dynamic soft cueing. The temporal embedding and text evidence are organized into a context and then input into a large language model that has been adapted and fine-tuned. This solves the problems of the failure to unify the modeling of temporal degradation features and health status semantics, as well as the insufficient stability of long-term prediction and the limited cross-domain generalization in scenarios with few samples, cross-operating conditions and cross-equipment scenarios. It achieves deep alignment between temporal features and text semantics, and improves the accuracy of remaining lifetime prediction and the stability of numerical output for complex equipment under changing operating conditions.

[0008] Based on the above scheme, the equipment remaining life prediction method of the present invention can be further improved as follows.

[0009] In one alternative approach, each sample window of the multivariate monitoring time series data is represented as: , For window length, To monitor the number of channels; the structured degradation text includes static and dynamic descriptions. The static description includes device type, data domain information, operating conditions and operating protocols. The dynamic description includes the loop position corresponding to the window endpoint, the target degradation label value and descriptors related to degradation.

[0010] The beneficial effects of adopting the above optional method are as follows: the sample window is further standardized as a matrix form containing the window length and the number of monitoring channels, and a structured degradation text covering static information such as equipment type and operating conditions as well as dynamic information such as cycle position and degradation label value is constructed, so that time series data and text description form a structured pair, providing a data foundation for cross-modal alignment.

[0011] In one alternative approach, the step of inputting each sample window into a timing encoder to obtain a timing representation includes: Each sample window is divided into multiple non-overlapping segments along the time axis; Map each non-overlapping segment to a segment embedding to obtain a segment embedding sequence; The temporal representation is obtained by extracting temporal features from the embedded sequence of the fragment using a multi-layer self-attention mechanism.

[0012] The advantages of adopting the above optional approach are as follows: the sample window is further divided into multiple non-overlapping segments along the time axis and mapped to segment embedding sequences. Then, temporal features are extracted through a multi-layer self-attention mechanism, enabling the temporal encoder to capture long-distance dependencies and local degradation patterns, thereby improving the ability of temporal representation to characterize the degradation process.

[0013] In one alternative approach, the temporal representation is projected onto a shared semantic space to obtain a projected temporal representation. The text representation is projected onto the shared semantic space to obtain the projected text representation. The similarity between the projection temporal representation and the projection text representation Represented as: in, For temperature coefficient, This is the similarity function.

[0014] The beneficial effects of adopting the above optional approach are as follows: the temporal representation and the text representation are further projected onto the shared semantic space, and the normalized similarity between the projected temporal representation and the projected text representation is calculated by a similarity function with a temperature coefficient, so that the features of the two modalities are comparable in a unified space, providing a stable metric basis for cross-modal comparative learning.

[0015] In one alternative approach, the contrastive loss includes a temporal-to-text direction loss. and text-to-temporal direction loss The time-to-text direction loss is expressed as: The text-to-temporal direction loss is expressed as: in, For batch size, For the first The projection timing representation and the first Similarity between projected text representations For the first The projection timing representation and the first Similarity between projected text representations For the first The projection timing representation and the first Similarity between projected text representations.

[0016] The beneficial effects of adopting the above optional approach are as follows: further designing contrast loss functions in both the temporal-to-text and text-to-temporal directions, respectively constraining the matching degree of positive sample pairs and suppressing negative sample interference, so that the temporal encoder can obtain symmetrical and consistent optimization objectives in the cross-modal contrastive pre-training process, thereby enhancing the alignment stability of the shared semantic space.

[0017] In one alternative approach, the remaining lifetime label range is the value range of continuous lifetime labels, and the ordered prototype value is obtained by uniformly sampling within the value range according to a preset sampling granularity; the text anchor is structured text containing the ordered prototype value, and the format of the structured text is a format indicating that the ordered prototype value is the remaining lifetime value.

[0018] The beneficial effects of adopting the above optional method are as follows: further, within the range of continuous lifetime label values, ordered prototype values ​​are uniformly sampled at a preset sampling granularity, and the prototype values ​​are converted into structured text anchors that indicate the remaining lifetime values, so that continuous lifetime labels and discrete semantic prototypes are established to establish a correspondence, providing a lifetime semantic reference benchmark with controllable granularity for large language models.

[0019] In one alternative approach, the target timing representation With the prototype semantic vectors Match score between Represented as: in, For learnable mapping matrix, For temperature parameters, This is an inner product operation; The fuzzy membership distribution The Middle element Represented as: in, The total number of the ordered prototype values.

[0020] The beneficial effects of adopting the above optional method are as follows: after projecting the target temporal representation with the learnable mapping matrix, the matching score is calculated with the prototype semantic vector, and then the fuzzy membership distribution is obtained by softmax operation, so that the degree of association between the sample window and multiple lifetime prototypes can be quantitatively expressed, providing a fine-grained semantic weight basis for dynamic soft prompt generation.

[0021] Secondly, the present invention provides a system for predicting the remaining life of equipment, the technical solution of which is as follows: The module is used to perform sliding segmentation on the multivariate monitoring time series data of the device to be predicted, obtain multiple sample windows, and construct the structured degraded text corresponding to each sample window; The training module is used to input each sample window into the temporal encoder to obtain a temporal representation, input each structured degenerate text into the pre-trained text encoder and map it to the shared semantic space through a projection layer to obtain a text representation, construct a contrastive loss based on the matched temporal-text pairs and the mismatched temporal-text pairs, perform cross-modal contrastive pre-training on the temporal encoder to obtain the pre-trained temporal encoder, and freeze the pre-trained text encoder. The calculation module is used to sample multiple ordered prototype values ​​in the remaining lifetime label interval and convert them into text anchors. The frozen pre-trained text encoder encodes each text anchor into a prototype semantic vector. For any sample window, the sample window is input into the pre-trained temporal encoder to obtain the corresponding target temporal representation. The matching score between the target temporal representation and each prototype semantic vector is calculated and the fuzzy membership distribution is obtained through the softmax function. The processing module is used to obtain dynamic soft prompts from the fuzzy membership distribution input prompt learner, project the target temporal representation into the temporal embedding of a large language model, and organize the dynamic soft prompts, the temporal embedding, textual evidence, and task instructions into a conditional input context. The prediction module is used to input the conditional input context into a large language model with efficient parameter fine-tuning, generate a structured response, parse the numerical fields in the structured response, and obtain the remaining lifetime prediction result of the device to be predicted.

[0022] The beneficial effects of the equipment remaining life prediction system of the present invention are as follows: The system of this invention performs sliding segmentation on multivariate monitoring time series and constructs structured degradation text. It uses cross-modal contrastive pre-training to map the temporal representation and text representation to a shared semantic space. It samples ordered prototype values ​​in the remaining lifetime label interval to obtain prototype semantic vectors. Combined with fuzzy membership distribution, it generates dynamic soft cueing. The temporal embedding and text evidence are organized into a context and then input into a large language model that has been adapted and fine-tuned. This solves the problems of the failure to unify the modeling of temporal degradation features and health status semantics, as well as the insufficient stability of long-term prediction and the limited cross-domain generalization in scenarios with few samples, cross-operating conditions and cross-equipment. It achieves deep alignment between temporal features and text semantics, and improves the accuracy of remaining lifetime prediction and the stability of numerical output for complex equipment under changing operating conditions.

[0023] Thirdly, the technical solution of an electronic device according to the present invention is as follows: The device includes a memory, a processor, and a program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps of the device remaining lifetime prediction method of the present invention.

[0024] Fourthly, the technical solution of a computer-readable storage medium provided by the present invention is as follows: The computer-readable storage medium stores instructions that, when read, cause the computer-readable storage medium to perform the steps of the device remaining lifetime prediction method of the present invention.

[0025] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other objects, features and advantages of the present invention more obvious and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0026] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a schematic flowchart of an embodiment of a method for predicting the remaining life of equipment according to the present invention; Figure 2 This is a schematic diagram of the overall framework; Figure 3 This is a schematic diagram of cross-modal semantic contrast pre-training. Figure 4 A diagram illustrating the construction of structured text descriptions; Figure 5 This is a schematic diagram of an embodiment of the equipment remaining life prediction system of the present invention; Figure 6 This is a schematic diagram of an embodiment of an electronic device according to the present invention. Detailed Implementation

[0027] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0028] Figure 1 This diagram illustrates a flowchart of an embodiment of a device remaining lifetime prediction method provided by the present invention. This method can be executed by an electronic device such as a terminal device or a server. The terminal device can be any fixed or mobile terminal, such as user equipment (UE), mobile device, user terminal, terminal, cellular phone, cordless phone, personal digital assistant (PDA), handheld device, computing device, in-vehicle device, or wearable device. The server can be a single server or a server cluster consisting of multiple servers. Any electronic device can implement the device remaining lifetime prediction method by having its processor call computer-readable instructions stored in its memory. Figure 1 As shown, it includes the following steps: S1. Perform sliding segmentation on the multivariate monitoring time series data of the device to be predicted to obtain multiple sample windows, and construct the structured degradation text corresponding to each sample window.

[0029] The device to be predicted refers to electromechanical equipment that requires remaining life assessment; for example, the device to be predicted is a type A battery undergoing cyclic charge-discharge testing. Multivariate monitoring time series data refers to a data sequence composed of measurements from multiple physical quantity channels collected in chronological order during equipment operation; for example, during the cyclic aging process of a type A battery, for each charge-discharge cycle, multiple degradation-related descriptors are extracted from monitoring data such as voltage and current, including constant-voltage charging capacity, constant-voltage charging duration, current slope, and current entropy, and these are arranged in cyclic order to form a multivariate monitoring time series.

[0030] Here, a sample window refers to a fixed-length time segment obtained by sliding the multivariate monitoring time series data according to a preset window length and step size. For example, the preset window length is 128 cycles, and the step size is 1 cycle. The first sample window includes the multivariate descriptors corresponding to the 1st to 128th cycles, the second sample window includes the multivariate descriptors corresponding to the 2nd to 129th cycles, and so on. Each sample window is represented as... ,in For window length, To monitor the number of channels; the first one in the window The observation vector at time t is denoted as , .

[0031] The structured degradation text refers to the natural language description text that organizes the static attribute information and dynamic degradation information of the device corresponding to the sample window according to the preset field order. For example, for the first sample window of the type A battery, the generated structured degradation text is "Device type: type A battery; Data field: laboratory aging data; Operating conditions: constant current charging and discharging, ambient temperature 25℃; Operating protocol: 1C (charge and discharge rate) constant current charging to 4.2V, 1C constant current discharging to 2.5V; Cycle position: 128th cycle; Current remaining life tag RUL: 58 cycles; Key descriptors - constant voltage charging capacity: 1.024Ah, constant voltage charging time: 18.5min, current slope: -0.023A / s, current entropy: 0.145".

[0032] It should be noted that the aforementioned structured degradation text is primarily used in the cross-modal comparison pre-training stage. The target degradation label value is the known label corresponding to the training sample, used to establish the cross-modal correspondence between temporal degradation features and health state semantics. In the subsequent few-shot adaptation and prediction stages, the target degradation label value is not used for the sample to be predicted. Instead, textual evidence is constructed based on the device attributes, operating conditions, degradation-related descriptors, and their changing trends corresponding to the sample. This textual evidence, along with temporal embeddings, dynamic soft prompts, and task instructions, serves as the conditional input to the large language model to generate the remaining lifetime prediction result.

[0033] S2. Input each sample window into the temporal encoder to obtain a temporal representation. Input each structured degenerate text into the pre-trained text encoder and map it to the shared semantic space through a projection layer to obtain a text representation. Contrast loss is constructed based on matching temporal-text pairs and unmatched temporal-text pairs. Cross-modal contrastive pre-training is performed on the temporal encoder to obtain the pre-trained temporal encoder. The pre-trained text encoder is then frozen.

[0034] The temporal encoder refers to a deep learning network used to map a multivariate time series sample window into a fixed-length vector representation. For example, the temporal encoder employs a segment-based transformer encoder, dividing a 128×3 dimension sample window along the time axis into 16 non-overlapping segments, each segment having a length of 8 time points. Each segment is mapped to a 256-dimensional segment embedding. These 16 segment embeddings form an embedding sequence, which is then processed through a multi-layer self-attention mechanism to obtain a temporal feature sequence. Window-level temporal features are obtained through aggregation labeling, pooling operations, or other sequence aggregation methods. Finally, a mapping layer yields a 512-dimensional vector as the temporal representation. The temporal encoder then processes the sample window... The encoding process is represented as ,in This is a timing encoder function. For its trainable parameters, This is the timing representation of the output.

[0035] The temporal representation refers to the fixed-length vector output by the temporal encoder after extracting features from the input sample window, which is used to characterize the degradation state information of the sample window. For example, for the first sample window of the type A battery, the temporal encoder outputs a 512-dimensional vector as the temporal representation of the window.

[0036] Here, a pre-trained text encoder refers to a language model pre-trained on a large-scale corpus to map natural language text into vector representations. For example, the pre-trained text encoder uses text encoder model A. After inputting structured degenerate text into text encoder model A, the specific tag vector output by the encoder is taken as the encoding result of the entire text. The pre-trained text encoder processes structured degenerate text... The encoding process is represented as ,in For pre-trained text encoder functions, This is the text encoding vector for the output.

[0037] In this context, the projection layer refers to a trainable linear transformation layer that maps the vector representation of one modality to the vector space of another modality. For example, the 768-dimensional text vector output by the pre-trained text encoder is mapped to a 512-dimensional shared semantic space through the projection layer. The projection layer is a linear transformation matrix with an input dimension of 768 and an output dimension of 512. The shared semantic space refers to the same-dimensional vector space in which the temporal representation and the text representation reside after their respective projection layers, ensuring comparability between the representations of different modalities. For example, both the temporal representation and the text representation reside in a 512-dimensional vector space after projection; this 512-dimensional vector space is the shared semantic space.

[0038] Here, text representation refers to the vector obtained by encoding structured degenerate text by a pre-trained text encoder and mapping it to a shared semantic space through a projection layer. For example, for the structured degenerate text of a type A battery, the pre-trained text encoder outputs a 768-dimensional text encoding vector, which is then mapped to a 512-dimensional text representation through a projection layer.

[0039] A matched time-text pair refers to a positive sample pair consisting of a time-series representation from the same sample window and the text representation of the corresponding structured degenerate text. For example, the time-series representation of the first sample window of the Type A battery and the text representation of the corresponding structured degenerate text constitute a matched time-text pair. A mismatched time-text pair refers to a negative sample pair consisting of a time-series representation from different sample windows and the text representation of the corresponding structured degenerate text. For example, the time-series representation of the first sample window of the Type A battery and the text representation of the corresponding structured degenerate text in the fifth sample window constitute a mismatched time-text pair.

[0040] Here, contrastive loss refers to a loss function used to constrain matching time-text pairs to be closer in the shared semantic space and mismatched time-text pairs to be farther apart. For example, in a training batch containing 32 sample windows, the similarity matrix between each time representation and all text representations is calculated, and the contrastive loss is constructed such that the similarity of matching time-text pairs is higher than that of mismatched time-text pairs.

[0041] Specifically, the temporal representation is projected onto the shared semantic space to obtain the projected temporal representation. Projecting the text representation onto a shared semantic space yields the projected text representation. Similarity between projected temporal representation and projected text representation Represented as ,in For temperature coefficient, This is the similarity function. The contrastive loss includes time-to-text direction loss. and text-to-temporal direction loss The time-to-text direction loss is expressed as The text-to-time direction loss is represented as ,in For batch size, For the first The projection timing representation and the first Similarity between projected text representations For the first The projection timing representation and the first Similarity between projected text representations For the first The projection timing representation and the first Similarity between projected text representations. Total loss from cross-modal semantic contrastive pre-training. The average of the losses in both directions is expressed as: .

[0042] The pre-trained temporal encoder refers to a temporal encoder whose parameters have been updated and optimized after cross-modal contrastive pre-training. For example, the temporal encoder is pre-trained on a dataset containing 100,000 time-text pairing samples to obtain the pre-trained temporal encoder. The pre-trained temporal encoder has the function of mapping the input sample window into a temporal representation with degenerate semantic awareness.

[0043] S3. Sample multiple ordered prototype values ​​in the remaining lifetime label interval and convert them into text anchors. Use the frozen pre-trained text encoder to encode each text anchor into a prototype semantic vector. For any sample window, input the sample window into the pre-trained temporal encoder to obtain the corresponding target temporal representation. Calculate the matching score between the target temporal representation and each prototype semantic vector, and obtain the fuzzy membership distribution through the softmax function.

[0044] The remaining life tag range refers to the range of the number of remaining cycles of the device from the current time to the time of failure; for example, the remaining life tag range of a type A battery is 0 to 100 cycles.

[0045] Here, ordered prototype values ​​refer to representative lifetime values ​​obtained by sampling within the remaining lifetime label range according to a preset sampling granularity and arranged in ascending order; for example, within the remaining lifetime label range of 0 to 100 times, 11 ordered prototype values ​​are obtained by sampling with a sampling granularity of 10, namely 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, and 100. Text anchors refer to natural language text converted from ordered prototype values ​​according to a preset structured format; for example, ordered prototype value 10 is converted to the text anchor "RUL=10", and ordered prototype value 50 is converted to "RUL=50".

[0046] Here, a frozen pre-trained text encoder refers to a pre-trained text encoder whose parameters remain unchanged and whose gradients are not updated during subsequent training. For example, after cross-modal contrastive pre-training, all parameters of the pre-trained text encoder are frozen and not updated in subsequent prototype semantic vector generation and downstream adaptation training. A prototype semantic vector refers to the encoded vector output by inputting text anchors into the frozen pre-trained text encoder. For example, inputting the text anchor "RUL=10" into the frozen text encoder model A and passing it through a projection layer outputs a 512-dimensional prototype semantic vector, with 11 ordered prototype values ​​corresponding to 11 512-dimensional prototype semantic vectors.

[0047] The target temporal representation refers to the temporal representation output by a pre-trained temporal encoder for any given sample window. For example, inputting the third sample window of a type A battery into the pre-trained temporal encoder yields a 512-dimensional target temporal representation. The matching score is the similarity value between the target temporal representation and each prototype semantic vector, used to measure the degree of matching between the sample window and the health state represented by each ordered prototype value. For example, the matching score between the target temporal representation and the prototype semantic vector "RUL=50" is 0.72, the matching score between it and "RUL=60" is 0.68, and the matching score between it and "RUL=40" is 0.55. Specifically, the target temporal representation... With the prototype semantic vectors Match score between Represented as ,in For learnable mapping matrix, For temperature parameters, This is for inner product operations.

[0048] The fuzzy membership distribution refers to the probability distribution obtained by normalizing the matching scores between the target temporal representation and all prototype semantic vectors using the softmax function. It is used to characterize the membership weight of a sample window relative to each ordered prototype value. For example, for 11 ordered prototype values, the fuzzy membership distribution of the third sample window of the type A battery is [0.01, 0.02, 0.03, 0.05, 0.08, 0.15, 0.25, 0.22, 0.11, 0.06, 0.02], with the highest membership weights at prototype values ​​RUL=60, RUL=50, and RUL=70. The Middle element Represented as ,in The total number of ordered prototype values. .

[0049] S4. The fuzzy membership distribution is input to the prompt learner to obtain dynamic soft prompts, the target temporal representation is projected into the temporal embedding of a large language model, and the dynamic soft prompts, the temporal embedding, textual evidence, and task instructions are organized into a conditional input context.

[0050] The cue learner refers to a trainable network that converts fuzzy membership distributions into dynamic soft cue vectors. For example, the cue learner consists of two one-dimensional convolutional layers, a ReLU nonlinear activation layer, a global average pooling layer, and two multilayer perceptron layers, converting an 11-dimensional fuzzy membership distribution into 16 dynamic soft cue vectors with a dimension of 768. Dynamic soft cues refer to conditional soft cue vectors generated by the cue learner based on the fuzzy membership distribution of the current sample window, used to guide the large language model to focus on the health state range of the current sample. For example, for the third sample window of a type A battery, the cue learner outputs 16 dynamic soft cue vectors with a dimension of 768, with the dynamic soft cue biased towards the health state range around RUL=60. Basic soft hints shared globally The final soft hint is obtained by superimposing the results. , represented as .

[0051] Temporal embedding refers to the vector sequence obtained by mapping the target temporal representation to the word embedding space of a large language model through a projection matrix. For example, a 512-dimensional target temporal representation is mapped to 16 temporal embedding vector sequences with a dimension of 768 through a linear projection matrix. Through projection matrix target time series representation The result is obtained by performing a linear transformation, and is expressed as follows: .

[0052] Here, textual evidence refers to the textual information consisting of the numerical values ​​and changing trends of key degradation descriptors extracted from the sample window; for example, for the third sample window of the type A battery, the textual evidence is "Current cycle position: 138th cycle; Constant voltage charging capacity: 0.987Ah; Constant voltage charging duration: 19.2min; Current slope: -0.035A / s; Current entropy: 0.162". Task instructions refer to natural language prompts used to inform the large language model of the current task objective; for example, the task instruction is "Please estimate the remaining lifetime of the current device based on the provided temporal embedding and health status prompts".

[0053] The conditional input context refers to the input sequence of the large language model, which is formed by organizing temporal embeddings, dynamic soft cues, textual evidence, and task instructions in a preset order. For example, for the third sample window of a type A battery, the conditional input context sequentially includes a temporal embedding start identifier, 16 temporal embedding vectors, 16 dynamic soft cue vectors, a lexical sequence of textual evidence, and a lexical sequence of task instructions. Represented as ,in For timing embedding start identifier, For timing embedding, This is the final soft suggestion. For textual evidence The word sequence, This is a task instruction.

[0054] S5. Input the conditions into the context into a large language model with efficient parameter fine-tuning, generate a structured response, parse the numerical fields in the structured response, and obtain the remaining lifetime prediction result of the device to be predicted.

[0055] Among them, the large language model with efficient parameter fine-tuning refers to the large language model after the backbone parameters of the large language model are frozen, and only a small number of trainable parameters are introduced into some network layers for adaptation training. For example, using the large language model B as the large language model, a low-rank matrix is ​​introduced into the attention projection layer for low-rank adaptation, all backbone parameters of the model are frozen, and only the low-rank matrix parameters participate in the training, thus obtaining the large language model with efficient parameter fine-tuning.

[0056] Here, the structured response refers to the text output generated by the large language model under the constraints of the conditional input context, which contains parsable numerical fields; for example, the structured response generated by the large language model after efficient parameter fine-tuning is "RUL=58". The numerical field refers to the numeric part of the structured response that represents the predicted remaining lifespan; for example, in the structured response "RUL=58", the number 58 is the numerical field.

[0057] The remaining lifetime prediction result refers to the remaining lifetime value of the device obtained after parsing the numerical field from the structured response; for example, parsing the numerical field 58 from the structured response "RUL=58" yields the remaining lifetime prediction result of the type A battery at the current moment as 58 cycles.

[0058] During the few-sample adaptation training phase, a joint loss function is used to optimize the model parameters. Word-level cross-entropy loss is employed. Constraining the consistency of the generated response and the target response on the word sequence; numerical loss. Using the symmetrical mean absolute percentage error definition, it is expressed as: ,in To predict the remaining lifetime value, This represents the actual remaining lifespan value. To smooth out the terms and prevent the denominator from being zero, the overall optimization objective is... ,in These are the weighting coefficients for numerical loss.

[0059] The technical solution of this embodiment solves the problems of inconsistent modeling of time series degradation features and health status semantics, insufficient stability of long-term prediction and limited cross-domain generalization under scenarios with few samples, cross-operating conditions and cross-equipment. It achieves deep alignment between time series features and text semantics, and improves the accuracy of remaining life prediction and the stability of numerical output for complex equipment under changing operating conditions.

[0060] In one alternative approach, each sample window of the multivariate monitoring time series data is represented as: , For window length, To monitor the number of channels; the structured degradation text includes static and dynamic descriptions. The static description includes device type, data domain information, operating conditions and operating protocols. The dynamic description includes the loop position corresponding to the window endpoint, the target degradation label value and descriptors related to degradation.

[0061] In the above-mentioned optional methods, the sample window is further represented as a matrix containing the window length and the number of monitoring channels, and a structured degradation text covering static information such as equipment type and operating conditions as well as dynamic information such as cycle position and degradation label value is constructed, so that time series data and text description form a structured pair, providing a data foundation for cross-modal alignment.

[0062] In one alternative approach, the step of inputting each sample window into a timing encoder to obtain a timing representation includes: Each sample window is divided into multiple non-overlapping segments along the time axis.

[0063] Non-overlapping segments refer to non-overlapping time segments after the sample window is divided along the time axis. For example, a sample window with a length of 128 time points is divided into 16 non-overlapping segments. The first segment contains data from time points 1 to 8, the second segment contains data from time points 9 to 16, and each segment contains voltage, current, and temperature data from 8 time points.

[0064] Each non-overlapping segment is mapped to a segment embedding to obtain a segment embedding sequence.

[0065] Here, the fragment embedding sequence refers to the embedding vector sequence formed by mapping each non-overlapping fragment to a fragment embedding and arranging them in chronological order; for example, 16 non-overlapping fragments are mapped to 16 256-dimensional fragment embeddings, and the embeddings are arranged in chronological order of the fragments to form a 16×256 fragment embedding sequence.

[0066] The temporal representation is obtained by extracting temporal features from the embedded sequence of the fragment using a multi-layer self-attention mechanism.

[0067] Among them, the multi-layer self-attention mechanism refers to a neural network structure that transforms the features of the input sequence through multiple stacked self-attention layers, in order to capture the long-range dependencies between different positions in the sequence. For example, a 4-layer converter encoder layer is used, each layer containing a multi-head self-attention sub-layer and a feedforward network sub-layer, to extract temporal features from the 16×256 segment embedding sequence and output a 16×256 temporal feature sequence.

[0068] In the above-mentioned optional approach, the sample window is further divided into multiple non-overlapping segments along the time axis and mapped to segment embedding sequences. Then, temporal features are extracted through a multi-layer self-attention mechanism, enabling the temporal encoder to capture long-distance dependencies and local degradation patterns, thereby improving the ability of temporal representation to characterize the degradation process.

[0069] In one alternative approach, the temporal representation is projected onto a shared semantic space to obtain a projected temporal representation. The text representation is projected onto the shared semantic space to obtain the projected text representation. The similarity between the projection temporal representation and the projection text representation Represented as: in, For temperature coefficient, This is the similarity function.

[0070] In the above-mentioned optional approach, the temporal representation and text representation are further projected onto a shared semantic space, and the normalized similarity between the projected temporal representation and the projected text representation is calculated by a similarity function with a temperature coefficient, so that the two modal features are comparable in a unified space, providing a stable metric basis for cross-modal comparative learning.

[0071] In one alternative approach, the contrastive loss includes a temporal-to-text direction loss. and text-to-temporal direction loss The time-to-text direction loss is expressed as: The text-to-temporal direction loss is expressed as: in, For batch size, For the first The projection timing representation and the first Similarity between projected text representations For the first The projection timing representation and the first Similarity between projected text representations For the first The projection timing representation and the first Similarity between projected text representations.

[0072] In the above-mentioned optional approaches, further design contrast loss functions in two directions, temporal to text and text to temporal, to constrain the matching degree of positive sample pairs and suppress the interference of negative samples from both directions, so that the temporal encoder can obtain a symmetrical and consistent optimization objective in the cross-modal contrastive pre-training process, thereby enhancing the alignment stability of the shared semantic space.

[0073] In one optional approach, the remaining lifetime label range is the value range of continuous lifetime labels, and the ordered prototype value is obtained by uniformly sampling within the value range according to a preset sampling granularity; the text anchor is structured text containing the ordered prototype value, and the format of the structured text is a format indicating that the ordered prototype value is the remaining lifetime value.

[0074] The value range refers to the continuous numerical range defined by the upper and lower boundaries of the remaining life tag interval; for example, the value range of the remaining life tag interval for a type A battery is 0 to 100 cycles.

[0075] The preset sampling granularity refers to the size of the interval between two adjacent sampling points when sampling ordered prototype values ​​within the range of values. For example, with 10 cycles as the preset sampling granularity, 11 ordered prototype values ​​of 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, and 100 are obtained by sampling within the range of values ​​from 0 to 100.

[0076] The format of the remaining lifetime value refers to the text form representing the remaining lifetime value in the text anchor; for example, the ordered prototype value is represented as a text anchor in the format "RUL=numerical value", such as "RUL=50".

[0077] In the above-mentioned optional methods, ordered prototype values ​​are further sampled uniformly within the range of continuous lifetime label values ​​at a preset sampling granularity, and the prototype values ​​are converted into structured text anchors that indicate the remaining lifetime values, so that continuous lifetime labels are associated with discrete semantic prototypes, providing a lifetime semantic reference benchmark with controllable granularity for large language models.

[0078] In one alternative approach, the target timing representation With the prototype semantic vectors Match score between Represented as: in, For learnable mapping matrix, For temperature parameters, This is for inner product operations.

[0079] The fuzzy membership distribution The Middle element Represented as: in, The total number of the ordered prototype values.

[0080] In the above-mentioned optional methods, the target temporal representation is further projected by a learnable mapping matrix and matched with the prototype semantic vector to calculate a matching score. Then, a fuzzy membership distribution is obtained through softmax operation, which allows the degree of association between the sample window and multiple lifetime prototypes to be quantitatively expressed, providing a fine-grained semantic weight basis for dynamic soft cue generation.

[0081] like Figure 2As shown, the overall process of the device remaining lifetime prediction method in this embodiment includes a cross-modal semantic pre-training stage, a fuzzy health semantic bridging stage, and a numerically-aware large language model adaptation and generation stage. Multi-source time-series data is processed by a time-series semantic encoder to obtain a time-series representation, and structured text descriptions are processed by a text encoder to obtain a text representation. The two representations are mapped to a shared semantic space through a projection layer to achieve initial alignment; this process is the core of the cross-modal semantic pre-training stage. In the fuzzy health semantic bridging stage, ordered prototype values ​​are sampled within continuous lifetime label intervals and converted into text anchors. A frozen pre-trained text encoder is used to encode the text anchors into prototype semantic vectors. For any sample window, the sample window is input into the pre-trained time-series encoder to obtain the target time-series representation. The matching score between the target time-series representation and each prototype semantic vector is calculated and normalized using a softmax function to obtain a fuzzy membership distribution. In the numerically-aware large language model adaptation and generation stage, the fuzzy membership distribution input cue learner is converted into dynamic soft cues. These dynamic soft cues are then superimposed with learnable base cue words to form the soft cue words. Simultaneously, the temporal representation is projected into the temporal embedding of the large language model. The temporal embedding, soft cue words, textual evidence cue words, and task instructions are organized into the conditional input context of the large language model. Under the constraints of this conditional input context, the large language model generates a structured response, which, after parsing, outputs the remaining life expectancy prediction result. During training, a numerically-aware training objective is used to optimize the model, enabling it to possess generative regression capabilities under limited sample conditions. The temporal semantic encoder, text encoder, health status prototype pool, cue learner, and large language model text embedding layer work collaboratively throughout the process to achieve end-to-end prediction from multi-source temporal data input to remaining life expectancy result output.

[0082] like Figure 3 As shown, the cross-modal semantic contrastive pre-training process processes multi-source temporal data and structured text descriptions separately. Multi-source temporal data is encoded by a temporal semantic encoder to obtain a temporal representation, while the structured text description is encoded by a text encoder and mapped through a projection layer to obtain a text representation. The temporal encoder employs a segment-based transformer encoder, which, for the input window... The window is divided into multiple non-overlapping segments along the time axis, and each segment is mapped to a segment embedding to obtain a segment embedding sequence. Then, a multi-layer self-attention mechanism is used to extract temporal features from the segment embedding sequence to obtain a temporal representation. The encoding process from the input window to the timing representation is represented as follows: ,in This is a timing encoder function. These are trainable parameters. On the text side, a pre-trained text encoder is used to process structured degenerate text. Encoding is performed, and the encoding process is represented as follows: ,in For pre-trained text encoder functions, The output text is encoded into a vector, and the encoded result is mapped to the same representation dimension as the temporal representation through a text-side projection layer to obtain the text representation. The temporal representation is then projected onto a shared semantic space to obtain the projected temporal representation. Projecting the text representation onto a shared semantic space yields the projected text representation. Similarity between projected temporal representation and projected text representation Represented as: in For temperature coefficient, The similarity function calculates the similarity score between each temporal representation and each text representation within a training batch. Matching temporal windows and text descriptions form positive sample pairs, while mismatched temporal windows and text descriptions form negative sample pairs. The contrastive loss includes a temporal-to-text direction loss. and text-to-temporal direction loss The time-to-text loss is expressed as: The text-to-time direction loss is represented as: in For batch size, For the first The projection timing representation and the first Similarity between projected text representations For the first The projection timing representation and the first Similarity between projected text representations For the first The projection timing representation and the first Similarity between projected text representations. Total loss from cross-modal semantic contrastive pre-training. The average of the losses in both directions is expressed as: By constraining the semantic alignment in the temporal-to-text direction and the text-to-temporal direction through bidirectional contrastive learning loss, positive sample pairs are closer together and negative sample pairs are farther apart in the shared semantic space. The temporal encoder thus obtains a degradation-aware representation capability constrained by the semantics of device attributes, operating conditions and degradation state.

[0083] like Figure 4As shown, the construction process of structured degradation text includes two branches: static information acquisition and dynamic information acquisition. The static information acquisition branch extracts device type, data domain information, operating conditions and protocols, and operating environment background information as static information elements. The dynamic information acquisition branch extracts time location information, target degradation labels and corresponding values, and obtains degradation descriptors as dynamic information elements. Then, through relevance filtering, it selects key descriptors highly correlated with the degradation state from the degradation descriptors. Static and dynamic information elements work together to construct paired text, organizing it into a unified format of structured degradation text according to a preset field order and outputting it. The structured degradation text includes two parts: static description and dynamic description. The static description includes at least one or more of device type, data domain information, operating conditions, and protocols. The dynamic description includes at least the loop position corresponding to the window endpoint, the target degradation label value, and a window-level descriptor highly correlated with the degradation. By generating structured degradation text using a unified template, samples from different device types and data domains have a consistent semantic field order, ensuring a stable pairing relationship between the time-series window and the text description.

[0084] After completing cross-modal semantic contrastive pre-training, a prototype-driven fuzzy health semantic bridging mechanism is constructed. Within the feasible lifetime label range, samples are obtained according to a preset granularity. Each of the ordered prototype values ​​is converted into a corresponding text anchor. A frozen pre-trained text encoder is then used to encode each text anchor into a prototype semantic vector, forming a healthy state prototype pool. For any sample to be predicted... The temporal representation is obtained through the pre-trained temporal encoder. Timing representation With the prototype semantic vectors Match score between Represented as: in For learnable mapping matrix, For temperature parameters, This is an inner product operation. The matching score is used to obtain the fuzzy health semantic distribution through the softmax function. The sample to be predicted is relative to the first... Membership weights of health state prototypes Represented as: The fuzzy health semantic distribution of the sample to be predicted, composed of the membership weights of each prototype. Represented as: Fuzzy membership distribution can simultaneously preserve the membership relationships of samples on multiple neighboring prototypes, and is used to express the sequential relationships, relative positions and local uncertainties in the continuous degradation process. This distribution is not directly used as the final lifetime prediction result, but as a healthy semantic condition for subsequent cue learning and generative regression of large language models.

[0085] In the cue learning and generative regression stages, the fuzzy health semantic distribution will be... The input cue learner, consisting of at least two layers of one-dimensional convolutions, non-linear mappings, pooling operations, and a multilayer perceptron, is used to transform fuzzy health semantic distributions into dynamic soft cues. At the same time, retain a set of globally shared basic hints. Used to represent the general context of the task. The final soft hint is derived from the base hint. With dynamic soft prompts Superimposed The construction method is represented as follows: Time sequence representation By mapping to the word embedding space of a large language model through a projection layer, temporal embeddings are obtained. The mapping process is represented as: in This is the projection matrix from time sequence to word embedding. The selected endpoint descriptor values ​​and their linear slopes within the window are written as textual evidence. and in conjunction with task instructions Construct the input context of the generative model The construction method is represented as follows: in For timing embedding start identifier, For timing embedding, This is the final soft suggestion. For textual evidence The word sequence, The task instruction is as follows. The large language model uses a low-rank adaptation (LoRA) parameter efficient fine-tuning method for few-sample adaptation, introducing low-rank incremental parameters only in the attention projection layer or linear layer. Under the conditional input context constraints, the model outputs structured response text, and the numerical fields in the response text are parsed to obtain the final remaining lifespan prediction result.

[0086] During the few-sample adaptation training phase, a joint loss function is used to optimize the model parameters. This includes word-level cross-entropy loss. Constrain the consistency between the generated response and the target response on the word sequence, and calculate the numerical loss. Using the symmetrical mean absolute percentage error definition, it is expressed as: in To predict the remaining lifetime value, This represents the actual remaining lifespan value. This is the smoothing term. The overall optimization objective is a weighted sum of the word-level cross-entropy loss and the numerical-level loss, expressed as: in The weights are the numerical loss coefficients. During the pre-training phase, the temporal encoder and text-side projection layer are updated, while the main body of the pre-trained text encoder is frozen. During the few-shot adaptation phase, the main body of the pre-trained temporal encoder and the backbone of the large language model are frozen, and only the prototype bridging mapping layer, temporal projection layer, cue learner, and LoRA adaptation parameters are updated to reduce the risk of overfitting under few-shot conditions, while retaining the general degenerate semantic knowledge obtained from cross-modal pre-training in the source domain.

[0087] During the deployment inference phase, the current multivariate monitoring window of the device under test is input, and textual evidence is constructed by combining the device attributes, operating conditions and key numerical descriptors corresponding to the window. Then, time-series coding, prototype matching, fuzzy health semantic distribution generation, soft cue construction and large language model structure generation are executed in sequence. Finally, the numerical fields in the generated response are parsed to obtain the remaining life prediction value.

[0088] Figure 5 A schematic diagram of an embodiment of a device remaining life prediction system 200 provided by the present invention is shown. For example... Figure 5 As shown, the remaining life prediction system 200 for this device includes: The construction module 201 is used to perform sliding segmentation on the multivariate monitoring time series data of the device to be predicted, obtain multiple sample windows, and construct the structured degenerate text corresponding to each sample window; Training module 202 is used to input each sample window into the temporal encoder to obtain a temporal representation, input each structured degenerate text into the pre-trained text encoder and map it to the shared semantic space through a projection layer to obtain a text representation, construct a contrastive loss based on the matched temporal-text pairs and the mismatched temporal-text pairs, perform cross-modal contrastive pre-training on the temporal encoder to obtain the pre-trained temporal encoder, and freeze the pre-trained text encoder. The calculation module 203 is used to sample multiple ordered prototype values ​​in the remaining lifetime label interval and convert them into text anchors. The frozen pre-trained text encoder encodes each text anchor into a prototype semantic vector. For any sample window, the sample window is input into the pre-trained temporal encoder to obtain the corresponding target temporal representation. The matching score between the target temporal representation and each prototype semantic vector is calculated and the fuzzy membership distribution is obtained through the softmax function. Processing module 204 is used to obtain dynamic soft prompts from the fuzzy membership distribution input prompt learner, project the target temporal representation into the temporal embedding of a large language model, and organize the dynamic soft prompts, the temporal embedding, textual evidence, and task instructions into a conditional input context. The prediction module 205 is used to input the conditional input context into a large language model with efficient parameter fine-tuning, generate a structured response, parse the numerical fields in the structured response, and obtain the remaining lifetime prediction result of the device to be predicted.

[0089] In one alternative approach, each sample window of the multivariate monitoring time series data is represented as: , For window length, To monitor the number of channels; the structured degradation text includes static and dynamic descriptions. The static description includes device type, data domain information, operating conditions and operating protocols. The dynamic description includes the loop position corresponding to the window endpoint, the target degradation label value and descriptors related to degradation.

[0090] In one alternative embodiment, the training module 202 is specifically used for: Each sample window is divided into multiple non-overlapping segments along the time axis; Map each non-overlapping segment to a segment embedding to obtain a segment embedding sequence; The temporal representation is obtained by extracting temporal features from the embedded sequence of the fragment using a multi-layer self-attention mechanism.

[0091] In one alternative approach, the temporal representation is projected onto a shared semantic space to obtain a projected temporal representation. The text representation is projected onto the shared semantic space to obtain the projected text representation. The similarity between the projection temporal representation and the projection text representation Represented as: in, For temperature coefficient, This is the similarity function.

[0092] In one alternative approach, the contrastive loss includes a temporal-to-text direction loss. and text-to-temporal direction loss The time-to-text direction loss is expressed as: The text-to-temporal direction loss is expressed as: in, For batch size, For the first The projection timing representation and the first Similarity between projected text representations For the first The projection timing representation and the first Similarity between projected text representations For the first The projection timing representation and the first Similarity between projected text representations.

[0093] In one optional approach, the remaining lifetime label range is the value range of continuous lifetime labels, and the ordered prototype value is obtained by uniformly sampling within the value range according to a preset sampling granularity; the text anchor is structured text containing the ordered prototype value, and the format of the structured text is a format indicating that the ordered prototype value is the remaining lifetime value.

[0094] In one alternative approach, the target timing representation With the prototype semantic vectors Match score between Represented as: in, For learnable mapping matrix, For temperature parameters, This is an inner product operation; The fuzzy membership distribution The Middle element Represented as: in, The total number of the ordered prototype values.

[0095] It should be noted that the beneficial effects of the equipment remaining life prediction system 200 provided in the above embodiments are the same as those of the equipment remaining life prediction method described above, and will not be repeated here. Furthermore, the system provided in the above embodiments is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the system can be divided into different functional modules according to the actual situation to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process is detailed in the method embodiments, and will not be repeated here.

[0096] The equipment remaining life prediction system 200 of the present invention may be a computer program (including program code) running on a computer device. For example, the equipment remaining life prediction system 200 of the present invention is an application software that can be used to execute the corresponding steps in the equipment remaining life prediction method of the present invention.

[0097] In some embodiments, the device remaining lifetime prediction system 200 of the present invention can be implemented in a combination of hardware and software. As an example, the device remaining lifetime prediction system 200 of the present invention can be a processor in the form of a hardware decoding processor, which is programmed to execute the device remaining lifetime prediction method of the present invention. For example, the processor in the form of a hardware decoding processor can be one or more application specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0098] The modules described in the embodiments of this invention can be implemented in software or hardware. The names of the modules are not, in some cases, limiting the scope of the module itself.

[0099] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the above-described device remaining lifetime prediction methods. That is, an electronic device according to an embodiment of the present invention may include, but is not limited to: a processor and a memory; the memory is used to store the computer program; the processor is used to execute the device remaining lifetime prediction method shown in any embodiment of the present invention by calling the computer program.

[0100] In one alternative embodiment, an electronic device is provided, such as Figure 6 As shown, Figure 6 The illustrated electronic device 4000 includes a processor 4001 and a memory 4003. The processor 4001 and the memory 4003 are connected, for example, via a bus 4002. Optionally, the electronic device 4000 may further include a transceiver 4004, which can be used for data interaction between the electronic device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver 4004 is not limited to one type, and the structure of the electronic device 4000 does not constitute a limitation on the embodiments of the present invention.

[0101] Processor 4001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this invention. Processor 4001 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0102] Bus 4002 may include a path for transmitting information between the aforementioned components. Bus 4002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 4002 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 6 The bus 4002 is represented by only one thick line, but this does not mean that there is only one bus or one type of bus.

[0103] The memory 4003 may be ROM (Read Only Memory) or other types of static storage devices capable of storing static information and instructions, RAM (Random Access Memory) or other types of dynamic storage devices capable of storing information and instructions, or EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0104] The memory 4003 stores application code (computer program) for executing the present invention, and its execution is controlled by the processor 4001. The processor 4001 executes the application code stored in the memory 4003 to implement the content shown in the foregoing method embodiments.

[0105] Among them, electronic devices can also be terminal devices. A terminal device can be any terminal device that can install applications and access web pages through applications, including at least one of smartphones, tablets, laptops, desktop computers, smart speakers, smartwatches, smart TVs, and smart in-vehicle devices.

[0106] It should be noted that, Figure 3 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0107] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-described methods for predicting the remaining lifespan of a device.

[0108] Alternatively, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, a floppy disk, and an optical data storage device, etc.

[0109] In an exemplary embodiment, a computer program product or computer program is also provided, which includes computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the electronic device to perform the aforementioned device remaining lifetime prediction method.

[0110] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0111] It should be understood that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0112] The computer-readable storage medium provided in this invention can be, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0113] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the method shown in the above embodiments.

[0114] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0115] It should be noted that the terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and represent a limitation on a specific order or sequence. Where appropriate, the order of use for similar objects can be interchanged so that the embodiments of this application described herein can be implemented in an order other than that shown or described.

[0116] Those skilled in the art will recognize that this invention can be implemented as a system, method, or computer program product. Therefore, this invention can be specifically implemented in the following forms: it can be entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, this invention can also be implemented as a computer program product contained in one or more computer-readable media, which includes computer-readable program code.

[0117] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for predicting the remaining life of equipment, characterized in that, include: The multivariate monitoring time series data of the device to be predicted is subjected to sliding segmentation to obtain multiple sample windows, and the structured degradation text corresponding to each sample window is constructed. Each sample window is input into the temporal encoder to obtain a temporal representation. Each structured degenerate text is input into the pre-trained text encoder and mapped to the shared semantic space through a projection layer to obtain a text representation. A contrastive loss is constructed based on the matched and unmatched temporal-text pairs. The temporal encoder is then pre-trained across modal contrastive modes to obtain the pre-trained temporal encoder. Finally, the pre-trained text encoder is frozen. Multiple ordered prototype values ​​are sampled in the remaining lifetime label interval and converted into text anchors. Each text anchor is encoded into a prototype semantic vector using the frozen pre-trained text encoder. For any sample window, the sample window is input into the pre-trained temporal encoder to obtain the corresponding target temporal representation. The matching score between the target temporal representation and each prototype semantic vector is calculated and obtained by the softmax function. The fuzzy membership distribution is input into the prompt learner to obtain dynamic soft prompts. The target temporal representation is projected into the temporal embedding of a large language model. The dynamic soft prompts, the temporal embedding, textual evidence, and task instructions are organized into a conditional input context. The conditions are input into the context of a large language model that has been efficiently fine-tuned with parameters to generate a structured response. The numerical fields in the structured response are then parsed to obtain the remaining lifetime prediction result of the device to be predicted.

2. The method for predicting the remaining life of equipment according to claim 1, characterized in that, Each sample window of the multivariate monitoring time series data is represented as follows: , For window length, To monitor the number of channels; the structured degradation text includes static and dynamic descriptions. The static description includes device type, data domain information, operating conditions and operating protocols. The dynamic description includes the loop position corresponding to the window endpoint, the target degradation label value and descriptors related to degradation.

3. The method for predicting the remaining life of equipment according to claim 2, characterized in that, The steps of inputting each sample window into the timing encoder to obtain the timing representation include: Each sample window is divided into multiple non-overlapping segments along the time axis; Map each non-overlapping segment to a segment embedding to obtain a segment embedding sequence; The temporal representation is obtained by extracting temporal features from the embedded sequence of the fragment using a multi-layer self-attention mechanism.

4. The method for predicting the remaining life of equipment according to claim 1, characterized in that, The time-series representation is projected onto the shared semantic space to obtain the projected time-series representation. The text representation is projected onto the shared semantic space to obtain the projected text representation. The similarity between the projection temporal representation and the projection text representation Represented as: in, For temperature coefficient, This is the similarity function.

5. The method for predicting the remaining life of equipment according to claim 4, characterized in that, The contrast loss includes time-to-text direction loss. and text-to-temporal direction loss The time-to-text direction loss is expressed as: The text-to-temporal direction loss is expressed as: in, For batch size, For the first The projection timing representation and the first Similarity between projected text representations For the first The projection timing representation and the first Similarity between projected text representations For the first The projection timing representation and the first Similarity between projected text representations.

6. The method for predicting the remaining life of equipment according to claim 1, characterized in that, The remaining lifetime label range is the value range of continuous lifetime labels, and the ordered prototype value is obtained by uniformly sampling within the value range according to a preset sampling granularity; the text anchor is structured text containing the ordered prototype value, and the format of the structured text is a format that indicates that the ordered prototype value is the remaining lifetime value.

7. The method for predicting the remaining life of equipment according to claim 6, characterized in that, The target time series representation With the prototype semantic vectors Match score between Represented as: in, For learnable mapping matrix, For temperature parameters, This is an inner product operation; The fuzzy membership distribution The Middle element Represented as: in, The total number of the ordered prototype values.

8. A system for predicting the remaining life of equipment, characterized in that, include: The module is used to perform sliding segmentation on the multivariate monitoring time series data of the device to be predicted, obtain multiple sample windows, and construct the structured degraded text corresponding to each sample window; The training module is used to input each sample window into the temporal encoder to obtain a temporal representation, input each structured degenerate text into the pre-trained text encoder and map it to the shared semantic space through a projection layer to obtain a text representation, construct a contrastive loss based on the matched temporal-text pairs and the mismatched temporal-text pairs, perform cross-modal contrastive pre-training on the temporal encoder to obtain the pre-trained temporal encoder, and freeze the pre-trained text encoder. The calculation module is used to sample multiple ordered prototype values ​​in the remaining lifetime label interval and convert them into text anchors. The frozen pre-trained text encoder encodes each text anchor into a prototype semantic vector. For any sample window, the sample window is input into the pre-trained temporal encoder to obtain the corresponding target temporal representation. The matching score between the target temporal representation and each prototype semantic vector is calculated and the fuzzy membership distribution is obtained through the softmax function. The processing module is used to obtain dynamic soft prompts from the fuzzy membership distribution input prompt learner, project the target temporal representation into the temporal embedding of a large language model, and organize the dynamic soft prompts, the temporal embedding, textual evidence, and task instructions into a conditional input context. The prediction module is used to input the conditional input context into a large language model with efficient parameter fine-tuning, generate a structured response, parse the numerical fields in the structured response, and obtain the remaining lifetime prediction result of the device to be predicted.

9. An electronic device, characterized in that, The electronic device includes a processor coupled to a memory storing at least one computer program, which is loaded and executed by the processor to enable the electronic device to implement the device remaining lifetime prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one computer program, which, when executed by a processor, implements the device remaining life prediction method as described in any one of claims 1 to 7.

Images