A device dynamic health degree fusion evaluation method based on multi-source time sequence data

Abstract

The application discloses a kind of equipment dynamic health degree fusion evaluation methods based on multi-source time series data, comprising: collecting and aligning working condition monitoring data and environmental monitoring data in equipment running cycle, and constructing multi-source time series training sample set with health label and environmental label;Introduce the causal representation model including environment encoding subnetwork and degradation encoding subnetwork, use the time window as contrast sample with the same health label but different environmental label, obtain the degradation representation with lower sensitivity to environmental change;Then input the degradation representation sequence into health degree regression network to generate dynamic health degree sequence, and set uniform health degree threshold on multiple devices and multiple working condition distribution, to give stable degradation degree evaluation results from real-time dynamic health degree sequence.

Description

Technical Field

[0001] This invention relates to the technical field of equipment condition monitoring and intelligent operation and maintenance, and in particular to a method for fusion assessment of equipment dynamic health based on multi-source time-series data. Background Technology

[0002] With the increasing digitalization of industrial equipment, power transmission and distribution equipment, and process control systems, a large amount of multi-source time-series monitoring data, including current, voltage, vibration, temperature, speed, load, and ambient temperature, humidity, and air pressure, is continuously generated throughout the equipment's life cycle. This field has gradually evolved from coarse-grained condition assessment based on single-point inspection and offline testing to dynamic health assessment based on online monitoring and data-driven models. The aim is to characterize the equipment degradation process through continuous health curves, supporting predictive maintenance and risk warning. However, in actual engineering scenarios, the health status of equipment is often strongly coupled with operating load and external environment. The same degradation level exhibits significantly different monitoring characteristics under different loads and environmental conditions, making the health assessment results highly sensitive to operating disturbances. Traditional methods usually use environmental and operating parameters as features input into a unified model, making it difficult to separate "degradation factors" and "environmental factors" at the representation level. Health curves lack comparability across equipment and scenarios, threshold settings rely on empirical adjustments, and false alarm and false negative rates are difficult to control, becoming a major bottleneck restricting the refinement and unification of equipment health management.

[0003] CN119782981A discloses a method and system for assessing the health status of multi-source heterogeneous power data. By constructing a unified data model for multi-source heterogeneous power data, the method performs time alignment, cleaning, and standardization of the heterogeneous data. Based on this, it extracts multi-dimensional features and assesses feature importance. Combining feature weighting and the degree of deviation from the benchmark value, it constructs a health status assessment model and further provides a health score curve that changes over time, enabling online monitoring and alarming of the operating status of power equipment. This solution represents a certain improvement in enhancing the fusion capability of multi-source heterogeneous data and providing dynamic health scores, but it does not address the environmental aspects. The method incorporates both quantity and operating condition quantity into a multi-dimensional feature set for weighted fusion, without introducing a distinction between health and environmental labels, nor setting independent representation channels for environmental and degradation factors in the model structure. As a result, the health score still fluctuates drastically with changes in operating conditions such as load and temperature, making it difficult to obtain comparable degradation health under different operating conditions. At the same time, the method relies more on feature contribution and fixed benchmark values ​​for threshold judgment, without constructing a unified threshold system from the perspective of health distribution across multiple devices and operating conditions. The threshold is strongly bound to the scene, which can easily lead to frequent adjustments of alarm thresholds and unstable recognition results under complex operating conditions.

[0004] CN113379182A discloses a method for assessing the health status of medium and low voltage equipment based on multidimensional state parameters. This method acquires multidimensional state parameters from monitoring data of medium and low voltage equipment, calculates the correlation between parameters using the Pearson correlation coefficient, constructs a membership function matrix and weight coefficients based on fuzzy hierarchical analysis, and integrates the assessment results of each sub-health status using DS evidence theory to obtain the final health status of the equipment. This method has strong engineering applicability in determining the weights of multiple indicators and fusing evidence, and can provide a relatively objective assessment of the equipment's health status from multiple perspectives. However, the overall evaluation process is somewhat static, focusing on outputting a single health level result at a given time segment, and does not construct a dynamic health status sequence that evolves continuously over time for multi-source time-series data. Furthermore, this method treats environmental factors as part of general state parameters, does not model the causal relationship between environmental conditions and equipment degradation, and lacks a mechanism to eliminate environmental disturbances at the representation level. Therefore, it is still difficult to support comparable assessments of different equipment and scenarios using a unified health status threshold under multiple operating conditions.

[0005] In summary, existing equipment health status assessment technologies based on multi-source data or multi-dimensional state parameters generally suffer from problems such as strong coupling between equipment degradation characteristics and environmental conditions, high sensitivity of dynamic health to environmental disturbances, and lack of unified health thresholds and comparable evaluation benchmarks across equipment and scenarios. This invention solves the problem of constructing a dynamic health assessment method for equipment based on multi-source time-series data that is insensitive to environmental disturbances and comparable across different equipment and operating conditions under multi-source time-series monitoring conditions. Summary of the Invention

[0006] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section, as well as in the abstract and title of the present application, to avoid obscuring the purpose of this section, the abstract and title of the invention. Such simplifications or omissions shall not be used to limit the scope of the present invention.

[0007] In view of the aforementioned existing problems, the present invention is proposed.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution: As a preferred embodiment of the device dynamic health fusion assessment method based on multi-source time series data described in this invention, the method involves: collecting operating condition monitoring data and environmental monitoring data during the device's operating cycle, aligning them with a unified timestamp and dividing them into time windows to obtain a multi-source time series training sample set with health and environmental labels. Based on the multi-source time-series training sample set, a causal representation model containing an environment coding sub-network and a degradation coding sub-network is constructed. Time window segments with the same health label but different environment labels are used as comparison sample pairs and input into the causal representation model to obtain a set of time window degradation representations in which the environment representation and degradation representation are decoupled and the degradation representation is less sensitive to environmental changes. The degradation representations of the time windows are concatenated in chronological order to form a degradation feature sequence and input into the health regression network. The health regression network is trained on time window segments with health labels to generate dynamic health sequences corresponding to each time window segment. Based on the dynamic health sequence, the health distribution of multiple devices under various operating conditions is statistically analyzed. A unified health threshold is set, and during the online monitoring phase, the real-time collected multi-source time-series data is sequentially input into the causal representation model and the health regression network to obtain the real-time dynamic health sequence, which is then compared with the unified health threshold to generate an assessment result of the device degradation degree.

[0009] The beneficial effects of this invention are as follows: In engineering scenarios where there are strong environmental disturbances and significant differences in equipment operating conditions in multi-source time-series data, this invention reduces the interference of environmental fluctuations on health assessment results, thereby reducing the risk of false alarms and missed alarms. On the other hand, it facilitates the adoption of unified operation and maintenance strategies and unified alarm standards in cross-unit, cross-site, and cross-process line scenarios, expanding from equipment-level health monitoring to consistent health management of group equipment, and improving the overall stability of equipment status assessment. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart illustrating the device dynamic health fusion assessment method based on multi-source time-series data as shown in this invention. Detailed Implementation

[0011] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0012] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort should fall within the scope of protection of this invention.

[0013] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0014] According to an embodiment of the present invention, in combination Figure 1 The flowchart shown illustrates a device dynamic health fusion assessment method based on multi-source time-series data, which specifically includes the following steps: S1. Collect operating condition monitoring data and environmental monitoring data during the equipment's operating cycle, and align and divide them into time windows according to a unified timestamp to obtain a multi-source time-series training sample set with health and environmental labels. Note that the following points should be noted in this step: S1.1 Collect operating condition monitoring data at a preset sampling interval. The operating condition monitoring data includes load, current, vibration, temperature and speed. Record a timestamp and equipment identifier for each operating condition monitoring data to obtain operating condition monitoring data with timestamps. In a preferred embodiment, a uniform sampling interval is set based on the control system clock during the equipment operation cycle. The sampling interval is set to 1 second. Acquisition channels are configured for load, current, vibration, temperature, and speed respectively. When data is acquired by each channel, the acquisition terminal automatically adds the current system time as a timestamp and adds the equipment identification code, so that the same equipment has a unique equipment identifier throughout the entire operation cycle. The raw data output by each sampling channel is written into the operating condition monitoring data table in ascending order of timestamp. The load, current, vibration, temperature, and speed acquired by the same equipment at the same timestamp are recorded as one line of operating condition monitoring data.

[0015] For example, with a sampling interval of 1 second, a single device can obtain 86,400 operational condition monitoring data entries with timestamps and device identifiers within 24 consecutive hours. The load, current, vibration, temperature, and speed fields in each record are fully filled, forming the basic operational condition data source for subsequent time sequence alignment and time window division.

[0016] S1.2 Collect environmental monitoring data at the time corresponding to the timestamp. The environmental monitoring data includes ambient temperature, humidity and air pressure. Record the timestamp and equipment identifier for each piece of environmental monitoring data. Correspond the environmental monitoring data with the operating condition monitoring data to obtain multi-source raw time series data. In a preferred embodiment, under the conditions of the same device identification and sampling interval as the working condition monitoring data, an ambient temperature sensor, a humidity sensor, and a barometric pressure sensor are arranged near the device installation environment, and sampling is triggered by a unified acquisition controller; when the controller receives a sampling instruction, it generates a time stamp for the environmental monitoring data according to the current system time, writes the ambient temperature, humidity, barometric pressure, time stamp, and device identification into the environmental monitoring data table together. Subsequently, using the time stamp and device identification as the association keys, the working condition monitoring data table and the environmental monitoring data table are associated, and the working condition monitoring data and environmental monitoring data corresponding to the same time stamp and the same device identification are combined into a record in the multi-source raw time series data.

[0017] Exemplarily, for the working condition monitoring data with a sampling interval of 1 s, the sampling interval of the environmental monitoring data is also set to 1 s. When there is a lack of environmental monitoring data at individual time stamps, a missing mark is first recorded and uniformly processed in step S1.3, thereby forming a multi-source raw time series data table with complete time stamp alignment keys and juxtaposed multi-source fields.

[0018] S1.3. Align the multi-source raw time series data according to the unified time stamp, and fill in the missing time stamps by interpolation to obtain multi-source time series aligned data synchronized on the same time axis; In a preferred embodiment, the multi-source raw time series data is grouped according to the device identification, sorted according to the time stamp within each device group to generate the time stamp sequence of the device. For the situation where there is an inconsistent sampling interval between the working condition monitoring data and the environmental monitoring data or individual records are missing, taking the time stamp sequence of the working condition monitoring data as the reference time axis, check whether there are complete records of ambient temperature, humidity, and barometric pressure at each time stamp on this time axis; if the environmental monitoring data is missing at a certain time stamp, take the environmental monitoring data collected at the adjacent front and rear time stamps of the device as the interpolation support points, and calculate the ambient temperature, humidity, and barometric pressure values of the missing time stamp according to the linear interpolation method, and add an interpolation identification field to the interpolation result.

[0019] Exemplarily, when two consecutive non-missing time stamps are t1 and t2, and the missing time stamp is t, and t1 < t < t2, the interpolated value of the ambient temperature is taken as the linear interpolation result of the temperatures at both ends, and the humidity and barometric pressure are filled in the same way, thereby obtaining multi-source time series aligned data synchronized on the same time axis.

[0020] S1.4. Divide the multi-source time series aligned data according to the preset time window length and sliding step size, combine the working condition monitoring data and environmental monitoring data within the same time window into the corresponding time window segment, and obtain multi-source time series segmented data containing multiple time window segments; In a preferred embodiment, the time window length and sliding step size are set according to the equipment degradation characteristic change cycle and operating condition fluctuation characteristics. For example, the time window length is set to 10 minutes and the sliding step size is set to 2 minutes. The sliding time window is constructed based on a unified timestamp sequence. For each device, in the aligned multi-source time series data, with the timestamp start point as t0, all operating condition monitoring data and environmental monitoring data within the time interval [t0, t0+10min) are included in the first time window segment. Then, the time start point is moved forward by 2 minutes to construct the second time window segment within the interval [t0+2min, t0+12min), until the entire operating cycle is covered.

[0021] For example, performing the above time window division on 24-hour aligned data can yield 715 time window segments. Each time window segment contains multiple records of load, current, vibration, temperature, rotational speed, and corresponding ambient temperature, humidity, and air pressure. These time window segments are aggregated to form multi-source time-series segmented data containing multiple time window segments.

[0022] S1.5. Label each time window segment with a health label based on the maintenance and fault records corresponding to the time window segment. Combine and divide the environmental temperature, humidity and load level into environmental labels according to the range of environmental monitoring data within each time window segment, thereby obtaining a multi-source time series training sample set with health labels and environmental labels.

[0023] Furthermore, a multi-source time-series training sample set with health and environmental labels is obtained. Specifically, this includes: matching maintenance and fault records based on the time range of each time window segment to determine the corresponding operating status record for each time window segment; dividing each time window segment into healthy, slightly degraded, and severely degraded states based on the operating status record, and writing the division results as health labels into each time window segment; statistically calculating the environmental monitoring data within each time window segment to obtain the time average values ​​and fluctuation ranges of environmental temperature, humidity, and load level, forming the corresponding environmental feature vector; using the environmental feature vector, determining the combination intervals of environmental temperature, humidity, and load level using preset segmentation rules and clustering results, and assigning environmental labels to time window segments falling into each combination interval; and extracting and summarizing each time window segment with health and environmental labels from the multi-source time-series segmented data to form a multi-source time-series training sample set with health and environmental labels.

[0024] In a preferred embodiment, the statistical method for environmental monitoring data within each time window segment is as follows: For each time window segment, extract the environmental temperature sequence, humidity sequence, and load level sequence corresponding to all sampling times within that time window segment, calculate the time average and fluctuation range for each, and combine the above statistical results into an environmental feature vector; assuming there are N sampling points within a certain time window segment, the corresponding environmental temperature sequence is... The humidity sequence is The load level sequence is Then the environmental feature vector e is expressed as: in, This represents the time-averaged ambient temperature within that time window segment. This represents the range of ambient temperature fluctuations within this time window segment; This represents the time-average humidity within that time window segment. This represents the range of humidity fluctuations within this time window segment; This represents the time average of the load level within this time window segment. This represents the range of load level fluctuations within this time window segment; This is the environmental feature vector corresponding to this time window segment.

[0025] For example, when the ambient temperature fluctuates between 20℃ and 25℃, the humidity fluctuates between 40% and 60%, and the load level changes between 0.6 and 0.8 within a certain time window, the average temperature, average humidity, and average load of that time window segment, as well as their respective fluctuation ranges, can be obtained through the above formula, thus forming an environmental feature vector that can be used for subsequent segmentation and clustering analysis.

[0026] In a preferred embodiment, the method for dividing the combined range of ambient temperature, humidity, and load level is as follows: Based on engineering experience of the equipment operating environment, a one-dimensional segmentation rule for ambient temperature, humidity, and load level is given. For example, the ambient temperature is divided into three segments: 0℃~15℃, 15℃~30℃, and above 30℃; the humidity is divided into three segments: 0%~40%, 40%~70%, and above 70%; and the load level is divided into three segments: 0~0.4, 0.4~0.7, and 0.7~1.0. This forms a set of temperature segments, a set of humidity segments, and a set of load level segments. The system first sets up segments; then, using the environmental feature vector obtained in S1.5.3 as input, it performs clustering analysis on all time window segments in the space of average temperature, average humidity, and average load. For example, it selects 5 to 8 cluster centers and records the one-dimensional segment combination in which each cluster center is located. Based on the one-dimensional segmentation rules and the distribution of cluster centers, it constructs environmental combination intervals, that is, it selects the segment combination with higher sample density in the three-dimensional segmentation space of temperature, humidity, and load level as candidate intervals, and assigns the same environmental label to the time window segments that fall into these candidate intervals.

[0027] For example, when a cluster center is located in a segmented combination area corresponding to an ambient temperature of 15℃~30℃, humidity of 40%~70%, and load level of 0.7~1.0, the area is defined as a high-load, medium-temperature and humidity environment range, and time window segments falling into this combination range are uniformly assigned the label of medium-temperature, medium-humidity, and high-load environment.

[0028] In a preferred embodiment, the classification of healthy state, mild degradation state, and severe degradation state is as follows: Fault marker field, alarm level field, and replacement component information field are extracted from the operating status record to establish a corresponding classification rule between the operating status record and the three types of healthy states. In the classification rule, records with fault shutdown markers and time window segments involving replacement records of key components such as spindle bearings, critical gearboxes, and main drive motors are classified as severely degraded state; time window segments containing only multiple alarm records or minor repair records but without fault shutdown events are classified as mildly degraded state; and time window segments with neither alarm records nor repair records are classified as healthy state.

[0029] Specifically, for each time window segment, check whether there is a fault shutdown mark or a critical component replacement record in its operating status record set. If so, write a severe degradation status in the health label field of the corresponding time window segment; if there are no such records but there are multiple alarm records or minor repair records, write a mild degradation status in the health label field; if there are neither alarm records nor any maintenance records, write a healthy status in the health label field.

[0030] For example, if a time window segment contains one alarm record and a spindle bearing replacement record appears in several subsequent time window segments, the aforementioned alarm time window segment can be marked as a slightly degraded state, while the time window segment containing the spindle bearing replacement record is marked as a severely degraded state, and the time window segment that does not overlap with any alarm or maintenance record is marked as a healthy state.

[0031] Preferably, this step addresses the technical problems in traditional equipment health assessments, such as the dispersed sources of operating condition data and environmental data, missing label information, and difficulty in comparing degradation processes under uniform time scales and environmental conditions. By establishing a clear data collection process, time alignment rules, time window segmentation parameter configuration, and joint labeling rules for health status and environmental labels, this step provides a well-structured and fully labeled training sample foundation for subsequent causal representation modeling and health regression training. This enables the acquisition of highly comparable degradation representations and dynamic health sequences under multiple equipment and operating conditions in subsequent steps, allowing for more refined differentiation and evaluation of equipment degradation processes in complex operating environments, and reducing the interference of environmental fluctuations on degradation identification results.

[0032] S2. Construct a causal representation model containing an environment coding sub-network and a degradation coding sub-network based on a multi-source time-series training sample set. Use time window segments with the same health label but different environment labels as comparison samples to input the causal representation model, resulting in a set of time window degradation representations where the environment representation and degradation representation are decoupled and the degradation representation has low sensitivity to environmental changes. Note that the following should be noted in this step: S2.1 Construct a causal representation model based on a multi-source time-series training sample set. Divide the causal representation model into an environment coding sub-network and a degenerate coding sub-network, and set up a shared input layer and independent coding layers for the causal representation model. In a preferred embodiment, each time window segment in the multi-source temporal training sample set obtained in S1 is regarded as a short sequence of length L, wherein each sampling point contains a dimension of L. The characteristics and dimensions of the working condition monitoring are as follows Environmental monitoring characteristics; for the same time window segment, all sampling points within that time window are arranged in chronological order, and the operating condition monitoring characteristics and environmental monitoring characteristics are concatenated along the feature dimension to form a dimension of The comprehensive feature vector, corresponding to the overall representation of the time window segment, is of dimension [dimensional value missing]. The input matrix is ​​used; the same time window length L and feature dimension d are used for all time window segments, thereby constructing a unified input tensor structure. , where N is the number of time window segments in the multi-source time-series training sample set; In a preferred embodiment, the method for constructing a shared input layer for a unified input tensor structure is as follows: each time window segment in the input tensor X is considered as a time series of length L. The comprehensive feature vector corresponding to each time step is first normalized by subtracting the mean and dividing by the standard deviation according to the feature dimension, ensuring that different physical quantities are within a comparable numerical range. Then, a temporal encoding vector is added to each time step in the time dimension. For example, a positional encoding method based on sine and cosine functions is used to map the time step index t to a fixed-length temporal embedding vector, which is then concatenated with the normalized comprehensive feature vector to form the shared feature representation output by the shared input layer. ,in To share feature dimensions; For example, preliminary feature extraction of the input tensor can be performed by single or multiple one-dimensional convolutions, gated recurrent units, or self-attention structures, so that the features output by the shared input layer simultaneously contain working condition and environmental information as well as temporal sequence information.

[0033] As an example, let H be the shared feature sequence output by the shared input layer, and let the environment coding subnetwork be the mapping function. The degenerate coding subnetwork is a mapping function. For a single time window segment input Corresponding shared features The environmental representation and degradation representation are as follows: in, To share the input layer parameter set, For the set of parameters of the environment coding sub-network, For the set of parameters of the degenerate coding subnetwork; This represents the environment corresponding to the k-th time window segment. This is the degenerate representation corresponding to the k-th time window segment; This represents the overall mapping of the causal representation model; In terms of network structure, both the environment coding subnetwork and the degenerate coding subnetwork adopt a multi-layer nonlinear mapping structure, such as a multi-layer feedforward network or a stacked timing module. The two only share the output of the shared input layer at the input end. It maintains independence in its internal coding layer and parameter set to separate environmental and degradation information under the same input conditions.

[0034] S2.2 Select time window segments with the same health label and different environmental label from the multi-source time-series training sample set, form a comparison sample pair with every two time window segments, and record the health label and environmental label of each comparison sample pair. In a preferred embodiment, in the multi-source time-series training sample set, all time window segments are first grouped according to health status, mild degradation status and severe degradation status based on health labels. Within each health status group, subgroups are formed according to environmental labels. Then, time window segments with different environmental labels are selected under the same health status and paired to form comparison sample pairs. The index, health label and environmental label of the two time window segments are saved in the comparison sample pair record. For example, when there are three environmental labels A, B, and C within the health status group, comparison sample pairs of three environmental combinations, AB, AC, and BC, can be constructed to ensure that time window segments under the same health status have sufficient differences in environmental conditions.

[0035] S2.3 Input the time window segments of each comparison sample pair into the environment coding sub-network to obtain the environment representation reflecting the changes in environmental temperature, humidity and load, and apply difference constraints to the environment representation in the same comparison sample pair; In a preferred embodiment, for each pair of comparison samples, let the environment representations obtained after inputting the two time window segments into the environment encoding subnetwork be respectively... and The difference between the two is measured using Euclidean distance, and a difference constraint term is introduced into the loss function: Where M is the number of comparison sample pairs. and These represent the environmental representations of the two time window segments in the i-th comparison sample pair; The difference constraint term, which is a Euclidean norm, has a negative coefficient in the overall loss function. When updating parameters, it drives the distance between the two environmental representations in the feature space to increase, thereby enhancing the ability of the environmental encoding sub-network to distinguish differences in environmental labels. It should be further noted that the comparison sample pairs can be divided into several training batches, and the above-mentioned difference constraint terms can be calculated in each training batch and weighted and combined with the degeneracy representation related constraints and regularization terms.

[0036] S2.4 Input the time window segments in each comparison sample pair into the degradation coding sub-network to obtain the degradation representation reflecting the degradation state of the device, and apply similarity constraints to degradation representations with the same health label and differentiation constraints to degradation representations with different health labels. In a preferred embodiment, the similarity and distinguishability constraints for the degenerate representation are as follows: Let For a set of pairs of time window segments with the same health label, For a set of different time window segments for health labels, For the sample pairs in the dataset, similarity constraints are constructed using Euclidean distance: for The sample pairs in the dataset are subjected to a comparison constraint with a gap: in, and This is a degenerate representation of the corresponding time window segment. and The number of similar and dissimilar pairs is denoted by m, which is the interval parameter. This similarity constraint and discrimination constraint simultaneously reduce the degradation representation distance of sample pairs with the same health label and expand the degradation representation distance of sample pairs with different health labels, thereby enabling the degradation coding subnetwork to have a higher resolution in distinguishing the degree of degradation.

[0037] S2.5. Construct a loss function based on difference constraints and similarity constraints, update the parameters of the causal representation model on the multi-source time series training sample set, and extract the degenerate representation corresponding to each time window segment from the degenerate representation output by the degenerate coding sub-network after training convergence, forming a set of time window degenerate representations with low sensitivity to environmental changes.

[0038] In a preferred embodiment, the environment representation difference constraint term, the degradation representation similarity constraint term, the distinguishability constraint term, and the parameter regularization term are combined into an overall loss function: in, , , and These are non-negative weighting coefficients. The set of all parameters of the causal representation model. The regularization term for the parameters is used during training to select comparison sample pairs in batches from the multi-source time-series training sample set and perform forward computation. and Within each batch, the loss value L is calculated, and then backpropagation is used to obtain the result. The gradient is calculated, and the parameters are updated using a gradient descent strategy; the training process is repeated in multiple rounds until the loss value fluctuates within the preset convergence range.

[0039] After training convergence, the causal representation model is input into each time window segment of the multi-source temporal training sample set, and the degenerate representation output by the degenerate coding sub-network is recorded. The data are then indexed and summarized according to time window segments to form a set of time window degradation representations that are less sensitive to environmental changes. This set serves as the input basis for the health regression network in subsequent S3.

[0040] S3. Concatenate the time window degradation representations in chronological order to form a degradation feature sequence and input it into the health regression network. Train the health regression network on time window segments with health labels to generate dynamic health sequences corresponding to each time window segment. Note that the following points should be noted in this step: S3.1 Sort the time window degradation representations according to the timestamps, and connect the time window degradation representations of the same device in chronological order to form a degradation feature sequence; In a preferred embodiment, the time window degradation representation set is first grouped according to the device identifier, and then arranged in ascending order according to the start timestamp of the time window within each device group to obtain the time window index sequence of the device. For each device group, the sorted time window degradation representations are connected sequentially to form the degradation feature sequence corresponding to the device. For cross-device training scenarios, when constructing batch input, a device-segmented approach is adopted to manage the degradation feature sequences of different devices in segments within the batch.

[0041] For example, for a certain device, its operating cycle contains more than 700 time window segments. After sorting the time window start time from early to late, they are recorded as the 1st to 700th time window degradation representations. These 700 degradation representations are connected in sequence to form a degradation feature sequence of length 700, so that the temporal evolution characteristics of the device degradation process can be reflected in the subsequent health regression network.

[0042] S3.2 Construct the input tensor of the health regression network based on the degradation feature sequence, and write the degradation representation of each time window and the corresponding time index into the input tensor; In a preferred embodiment, within each device group, based on the degradation feature sequence obtained in chronological order, the degradation representation of each time window is considered as having a dimension of The degenerate feature vector is constructed, and a time index vector is introduced to represent the time order. For each device, a time index t is assigned to the degenerate representation of the t-th time window in its degenerate feature sequence. The degenerate feature vector and the time index are combined into an extended degenerate feature vector by concatenation or embedding. Subsequently, in the batch construction stage, the degenerate feature sequences of several devices are aligned and padded to a uniform length, and the insufficient parts are padded with mask marks to form a dimension of 1. The input tensor is B, where B is the number of devices in the batch. The length of the longest degenerate feature sequence in the batch. This refers to the expanded feature dimensions.

[0043] For example, the input tensor can be constructed using a fixed-length sliding sequence method. The degenerative feature sequence of length 700 is divided into segments of length 50 by sliding windows. Each segment forms a local degenerative sequence fragment, and the corresponding time index interval is recorded in the input tensor. This preserves the local dynamics of the degenerative process while limiting the length of the network input.

[0044] S3.3. The time window with health label is degenerate and input into the health regression network via the input tensor. The regression loss function is used to constrain the deviation between the health output and the health label, and the parameters of the health regression network are updated. The process involves: writing the degenerate representation of a time window with a health label and its corresponding health label into the training batch input tensor, maintaining a one-to-one correspondence between the time window degenerate representation and the health label within the training batch input tensor; inputting the training batch input tensor into the health regression network to obtain the health output corresponding to each time window degenerate representation, and recording the pairing results of the health output and health label within the same batch; constructing a regression loss function based on the difference between the health output and health label corresponding to each time window degenerate representation, assigning higher weights to samples in mild and severe degenerate states and lower weights to samples in a healthy state, thus obtaining a weighted regression loss value; performing backpropagation on the health regression network with the weighted regression loss value as the objective, calculating the gradient of the network parameters at each layer, and updating the health regression network parameters using a gradient descent strategy at a preset learning rate; repeating the above steps in multiple training rounds to iterate through all time window degenerate representations with health labels until the weighted regression loss value stabilizes within a preset convergence range; In a preferred embodiment, the degradation of a single time window is represented as The corresponding health label is The health regression network is a mapping function. Then the health output for the k-th time window segment is expressed as: in, For the set of parameters of the health regression network, Output the health status corresponding to the k-th time window segment; To introduce importance weights for different health states during the training phase, healthy state, mildly deteriorated state, and severely deteriorated state are mapped to target health label values, respectively. For example, a corresponding health status can be set. Mild degradation corresponds to Severe degradation state corresponds to And set weight coefficients for different states. If the weights of healthy state samples are lower, and the weights of samples in mild and severe deterioration states are higher, then the weighted regression loss function within a single training batch is expressed as: Where B represents the number of time window segments within the current training batch. The weight coefficient for the k-th time window segment; The health score output by the health regression network for this time window segment. The target health label value is obtained by converting the health label. This represents the weighted regression loss value for the current training batch. In terms of network structure, the health regression network can adopt structures such as multi-layer feedforward networks, temporal convolutional networks, or recurrent neural networks. It outputs the corresponding scalar health value for each time window degradation representation in the input degradation feature sequence.

[0045] S3.4. In the training process of the health regression network, a health difference regularization term for adjacent time windows is introduced to ensure that the health output in the same degradation process remains continuous along the time direction and shows an overall downward trend. In a preferred embodiment, to introduce differential constraints on the health output of adjacent time windows during the degradation process, for each device's degradation feature sequence, let the number of time windows for that device be T, and the corresponding health output sequence be... To ensure that the degradation process exhibits a reasonable and slow downward trend over time, a differential regularization term is introduced for the health difference between adjacent time windows: Wherein, the summation index d represents different devices. Let d be the number of time windows for the d-th device. Output the health status of the d-th device in the t-th time window; and It is a non-negative weighting coefficient used to adjust the constraint strength on the upward trend of health and the magnitude of change in health. This is the overall loss value of the difference regularization term; the first part of the above regularization term Introducing a penalty for the rate at which health increases over time, Part Two. A smoothing constraint is introduced for the magnitude of changes in health status.

[0046] During the training phase, the weighted regression loss term and the difference regularization term are combined as the training objective of the health regression network: in, These are the weighting coefficients for the difference regularization term. The overall training loss value of the regression network is used to determine the health status.

[0047] For example, a larger value is set for equipment groups with more obvious degradation processes. and This makes the health output more closely resemble a gradual decline over time during the degradation phase, while setting a smaller value for equipment groups with more long-term healthy states. and This reduces the suppression of subtle fluctuations during the stable phase.

[0048] S3.5 After the health regression network is trained, the degradation representation of each time window is input into the health regression network in chronological order to obtain a dynamic health sequence that corresponds one-to-one with each time window segment.

[0049] In a preferred embodiment, after the health regression network is trained, for each device, the degenerate representation of all time windows of the device is input into the health regression network in chronological order, and the health values ​​output by the network are read in sequence. These health values ​​are paired with the corresponding time window segments to form the dynamic health sequence of the device. In cross-device scenarios, the dynamic health sequences of all devices and their corresponding health status labels are stored centrally, and the device identifier, the start and end time of the time window, and the health value are retained in the data structure.

[0050] For example, for a certain device, its operating cycle contains 700 time window segments. The health regression network outputs 700 scalar health values ​​for the degradation representation of these 700 time windows. These 700 health values ​​are arranged in chronological order to form the dynamic health sequence of the device, which provides the basis for the subsequent health threshold statistics and online evaluation in S4.

[0051] S4. Based on the dynamic health status sequence, statistically analyze the health status distribution of multiple devices under various operating conditions, set a unified health status threshold, and during the online monitoring phase, sequentially input the real-time collected multi-source time-series data into the causal representation model and the health status regression network to obtain the real-time dynamic health status sequence, compare it with the unified health status threshold, and generate the equipment degradation degree assessment result. Note that the following points should be noted in this step: S4.1 Based on the dynamic health status sequence of multiple devices and various working conditions, the health status values ​​corresponding to the healthy state, mild degradation state, and severe degradation state are statistically analyzed to obtain the health status distribution range of each state, and a candidate set of health status thresholds is constructed. In a preferred embodiment, based on the dynamic health status sequences obtained from multiple devices and various operating conditions, all time window segments are first divided into a healthy state set, a mildly degraded state set, and a severely degraded state set according to health tags. Within each set, the distribution of corresponding health status values ​​is statistically analyzed. Specifically, for the healthy state set, the common value ranges of the health status are determined by calculating the quantiles, mean, and standard deviation of the health status values. For the mildly degraded and severely degraded state sets, the low quantiles and concentration intervals of the health status values ​​are statistically analyzed respectively. Using the boundary region between the concentration intervals of health status values ​​for mildly degraded and severely degraded states and the concentration interval of health status values ​​for the healthy state as a reference, several candidate threshold points are constructed. For example, several candidate values ​​are selected near the boundary of the health status distribution intervals for the healthy and mildly degraded states, and additional candidate values ​​are added near the boundary of the health status distribution intervals for the mildly degraded and severely degraded states, thereby forming a candidate set of health status thresholds covering multiple devices under different operating conditions.

[0052] For example, when the health status of the healthy samples is concentrated in the range of 0.7 to 1.0, the mild deterioration status is concentrated in the range of 0.4 to 0.7, and the severe deterioration status is concentrated in the range of 0.0 to 0.4, several candidate threshold points can be selected in the range of 0.4 to 0.7 with a fixed step size, such as 0.45, 0.5, 0.55, 0.6, 0.65, etc., and these candidate thresholds are evaluated one by one in subsequent steps.

[0053] S4.2 Calculate the false alarm rate and false negative rate on historical samples of each device under each operating condition based on the candidate set of health thresholds, and select a unified health threshold that achieves a trade-off between the false alarm rate and false negative rate and has a small change in the identification results under each operating condition. In a preferred embodiment, for each candidate threshold The recognition performance of each device under various operating conditions was statistically analyzed using historical samples. Let H be the set of healthy state samples across all time windows, and D be the set of degraded state samples (the union of mild and severe degraded states). The health score of the k-th time window in the health score sequence is... ; Regarding candidate thresholds On historical samples, the false alarm rate and the false negative rate can be defined as follows: in, The proportion of healthy samples that are classified as degraded is the candidate threshold. The false alarm rate is below; The proportion of samples in a degraded state that are not classified as degraded; this is the candidate threshold. The false negative rate is below; This represents the number of samples in the set. To balance the false alarm rate and the false negative rate, and to consider the stability of the identification results under different operating conditions, the corresponding false alarm rate and false negative rate can be calculated separately under different operating conditions. Then, the variance of the identification index for a certain candidate threshold under different operating conditions is calculated to construct a comprehensive evaluation function. in, The coefficients used to control the weighting of false alarm rate and false negative rate; To control the coefficients for identifying stability weights under different operating conditions; Candidate threshold under operating condition c Accuracy of identifying historical samples; To improve the recognition accuracy under various operating conditions c The smaller the variance, the more stable the recognition results are under different working conditions; Candidate threshold Overall comprehensive evaluation value.

[0054] During the threshold selection phase, a search is performed in the candidate threshold set to make... Threshold for taking the minimum value This serves as a unified health threshold.

[0055] For example, when calculating the false alarm rate, false negative rate, and variance of recognition accuracy under different operating conditions for candidate thresholds of 0.45, 0.5, 0.55, and 0.6, respectively, if the threshold of 0.5 corresponds to... and The sum is small, and If the threshold is lower than other thresholds, 0.5 can be selected as the uniform health threshold.

[0056] S4.3 During the online monitoring phase, collect real-time multi-source time-series data, process the real-time multi-source time-series data according to timestamp alignment and time window division rules to obtain real-time time window segments, input the real-time time window segments into the causal representation model, and output the real-time time window degradation representation. In a preferred embodiment, the real-time multi-source time-series data collected during the online monitoring phase is processed using the same timestamp alignment and time window division rules as S1. For each device, the real-time collected operating condition monitoring data and environmental monitoring data are associated by timestamp, and real-time multi-source time-series aligned data is constructed on a unified time axis. Real-time time window segments are generated according to the preset time window length and sliding step size. Then, each real-time time window segment is used as input and sequentially passed through the shared input layer and degenerate coding sub-network of the trained causal representation model to obtain the corresponding real-time time window degenerate representation.

[0057] Let the real-time multi-source time window segment be... The shared input layer is mapped as Degenerate coding subnetwork is Then the degradation representation of the k-th real-time time window segment is: in, and These are the parameter sets of the shared input layer and the degenerate coding subnetwork obtained during the offline training phase, respectively. This is the degenerate representation of the real-time window corresponding to the k-th real-time time window segment.

[0058] S4.4 Connect the real-time time window degradation representations in chronological order to form a real-time degradation feature sequence, and input the real-time degradation feature sequence into the health regression network to obtain the real-time dynamic health sequence corresponding to each real-time time window segment; In a preferred embodiment, the real-time time window degradation representations obtained in S4.3 are concatenated in chronological order to form a real-time degradation feature sequence. Let the number of real-time time windows currently generated by a certain device during the online monitoring phase be... The corresponding degenerate representation sequence is The degenerate representation sequence is then sequentially input into the trained health regression network. For each time window, the degradation representation outputs the corresponding real-time health value. The real-time dynamic health sequence is as follows: in, This is the set of health regression network parameters obtained during the offline training phase. The real-time health value corresponding to the t-th real-time time window segment; Let be the degenerate representation vector of the t-th real-time time window segment; This represents the cumulative number of real-time time windows during the current online monitoring phase.

[0059] As online monitoring progresses, new real-time time window degradation representations are continuously added to the degradation feature sequence. The health regression network can output new health values ​​for the newly added degradation representations, thereby gradually extending the real-time dynamic health sequence.

[0060] For example, when a device runs continuously for 3 hours during the online monitoring phase, dozens of real-time time window segments can be obtained by dividing the time window into 10-minute time window lengths and 2-minute sliding steps. After generating a degradation representation for these time window segments, the health regression network outputs a real-time dynamic health value sequence, and the operator can observe the trend of the device's health over time on the time axis.

[0061] S4.5. Compare the health status of each time window in the real-time dynamic health status sequence with a unified health status threshold, and determine the healthy status, mild deterioration status and severe deterioration status based on the duration of continuous decline below the unified health status threshold and the rate of health status decline.

[0062] In a preferred embodiment, the comparison process between the health status of each time window in the real-time dynamic health status sequence and a unified health status threshold can be formalized as follows: Let the unified health status threshold be... The real-time dynamic health sequence is Then, high and low markers are defined for each time window: in, The health level of the t-th time window is marked as high or low relative to a unified health threshold. When the health level is lower than the threshold, it is marked as 1, and when the health level is not lower than the threshold, it is marked as 0. A continuous 1 segment in the above high and low labeled sequence is identified as a time period continuously below a uniform health threshold. Let the start and end time window index of a certain continuous segment be denoted as . and And within this interval, for all t, For this continuous time period, the duration of the continuous low threshold is calculated by converting the time window length and time window interval: in, The time interval between the start times of adjacent time windows. This represents the duration of the continuous time period.

[0063] Within the same time period, the difference in health status between different time windows is calculated sequentially based on the output of health status. The rate of health status decline is then calculated based on the time interval. For example, the average rate of decline can be defined as: in, This represents the average rate of change of health over the continuous time period. A negative value indicates that health decreases over time, and a larger absolute value indicates a faster rate of decrease.

[0064] In the state partitioning rules, a preset time length threshold is used. With the descent rate threshold (If negative), within each consecutive time period according to and The combination of these factors determines the state flag of the corresponding time window segment; an example rule can be expressed as: like and If so, the corresponding time window is marked as a state of mild degradation. like or If so, the corresponding time window is marked as severely degraded. For time windows that are not included in any consecutive low threshold time period, i.e. Furthermore, for time windows that are not included in the degradation period, the health status can be written in the status flag.

[0065] For example, when If the duration of a certain continuous low threshold period is only 10 minutes, and the average rate of decrease is... If the corresponding time window segment is marked as a slightly degraded state, then if the duration of another continuous time period is 30 minutes, the average rate of decline is... If so, the corresponding time window segment is marked as severely degraded.

[0066] Preferably, this embodiment constructs a multi-source time-series training sample set with health and environmental labels under a unified time axis, and introduces a causal representation model including an environmental encoding sub-network and a degradation encoding sub-network. This separates equipment degradation features from environmental operating condition features in the representation space, ensuring that the degradation representation remains relatively stable under different operating conditions such as high load, high temperature, and low temperature. Based on this, the time-window degradation representation is input into the health regression network in chronological order to obtain a dynamic health sequence that reflects the degree, trend, and rate of change of degradation. This transforms the equipment status from discrete fault / maintenance events. To provide a continuously quantified evolution trajectory, further, by statistically analyzing the dynamic health distribution across multiple devices and operating conditions and setting a unified health threshold, the degradation assessment under different devices and operating conditions is conducted based on the same health scale. In the online phase, the degradation assessment result is given directly based on the relationship between the real-time dynamic health sequence and the unified health threshold. Thus, in engineering scenarios where there are strong environmental disturbances in multi-source time-series data and significant differences in equipment operating conditions, the interference of environmental fluctuations on the health assessment results is reduced, the risk of false alarms and false negatives is decreased, and the overall stability of equipment status assessment is improved.

[0067] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for fusion assessment of device dynamic health based on multi-source time-series data, characterized in that, include: During the equipment's operating cycle, operational condition monitoring data and environmental monitoring data are collected, and these data are aligned and divided into time windows according to a unified timestamp to obtain a multi-source time-series training sample set with health and environmental labels. Based on the multi-source time-series training sample set, a causal representation model containing an environment coding sub-network and a degradation coding sub-network is constructed. Time window segments with the same health label but different environment labels are used as comparison sample pairs and input into the causal representation model to obtain a set of time window degradation representations in which the environment representation and degradation representation are decoupled and the degradation representation is less sensitive to environmental changes. Obtaining the time window degradation representation set includes: constructing a causal representation model based on the multi-source time-series training sample set; dividing the causal representation model into an environment encoding sub-network and a degradation encoding sub-network; and setting a shared input layer and independent encoding layers for the causal representation model; selecting time window segments with the same health label but different environment labels from the multi-source time-series training sample set; forming comparison sample pairs from every two time window segments; and recording the health label and environment label of each comparison sample pair; inputting the time window segments from each comparison sample pair into the environment encoding sub-network to obtain an environment representation reflecting changes in environmental temperature, humidity, and load; and performing the same... A difference constraint is applied to the environmental representation in each comparison sample pair; the time window segments in each comparison sample pair are input into the degradation encoding sub-network to obtain a degradation representation reflecting the degradation state of the device, and a similarity constraint is applied to degradation representations with the same health label, and a distinction constraint is applied to degradation representations with different health labels; a loss function is constructed based on the difference constraint and the similarity constraint, and the parameters of the causal representation model are updated on the multi-source time-series training sample set; after training convergence, the degradation representation corresponding to each time window segment is extracted from the degradation representation output by the degradation encoding sub-network to form a set of time window degradation representations with low sensitivity to environmental changes; The causal representation model is constructed as follows: The input feature dimension and time window length are determined based on the multi-source time-series training sample set; a unified input tensor structure is constructed for each time window segment; a shared input layer is established based on the input tensor structure; the working condition monitoring data and the environmental monitoring data are normalized and time-series encoded to obtain a shared feature representation; an environmental coding sub-network is set up based on the shared feature representation; multi-layer nonlinear mapping is performed on the environmental features in the shared feature representation to obtain an environmental coding output; a degradation coding sub-network is set up based on the shared feature representation; multi-layer nonlinear mapping is performed on the degradation features in the shared feature representation to obtain a degradation coding output; independent coding layers and independent parameter sets are set up in the environmental coding sub-network and the degradation coding sub-network respectively; information is limited between the two branches to be transmitted only through the shared input layer, so as to separate environmental information and degradation information under the same input conditions, thus constituting the causal representation model. The degradation representations of the time windows are concatenated in chronological order to form a degradation feature sequence and input into the health regression network. The health regression network is trained on time window segments with health labels to generate dynamic health sequences corresponding to each time window segment. Based on the dynamic health sequence, the health distribution of multiple devices under various operating conditions is statistically analyzed. A unified health threshold is set, and during the online monitoring phase, the real-time collected multi-source time-series data is sequentially input into the causal representation model and the health regression network to obtain the real-time dynamic health sequence, which is then compared with the unified health threshold to generate an assessment result of the device degradation degree.

2. The device dynamic health fusion assessment method based on multi-source time-series data according to claim 1, characterized in that, The multi-source time-series training sample set is obtained, including: Operating condition monitoring data is collected at a preset sampling interval. The operating condition monitoring data includes load, current, vibration, temperature, and speed. A timestamp and equipment identifier are recorded for each operating condition monitoring data to obtain operating condition monitoring data with timestamps. Environmental monitoring data is collected at the time corresponding to the timestamp. The environmental monitoring data includes ambient temperature, humidity, and air pressure. A timestamp and equipment identifier are recorded for each piece of environmental monitoring data. The environmental monitoring data is correlated with the operating condition monitoring data to obtain multi-source raw time-series data. The multi-source original time series data are aligned according to a unified timestamp, and missing timestamps are filled by interpolation to obtain multi-source time series aligned data synchronized on the same time axis. The multi-source time-series aligned data is divided into time windows according to a preset time window length and sliding step size. The working condition monitoring data and environmental monitoring data within the same time window are combined into corresponding time window segments to obtain multi-source time-series segmented data containing multiple time window segments. Health labels are assigned to each time window segment based on the maintenance and fault records corresponding to the time window segment. Environmental labels are then assigned by combining environmental temperature, humidity, and load level according to the range of environmental monitoring data within each time window segment, thereby obtaining a multi-source time-series training sample set with health and environmental labels.

3. The device dynamic health fusion assessment method based on multi-source time-series data according to claim 2, characterized in that, The obtained multi-source time-series training sample set with health and environmental labels specifically includes: Match the maintenance records and fault records according to the time range of each time window segment to determine the corresponding operating status record for each time window segment; Based on the running status record, each time window segment is divided into a healthy state, a mildly degraded state, and a severely degraded state, and the division result is written as a health label into each time window segment; Statistical calculations were performed on the environmental monitoring data within each time window segment to obtain the time average values ​​and fluctuation ranges of environmental temperature, humidity, and load level, thus forming the corresponding environmental feature vectors. Based on the environmental feature vector, a combination range of environmental temperature, humidity and load level is determined by using a preset segmentation rule and clustering results, and environmental labels are assigned to time window segments that fall into each combination range. Each time window segment with health and environmental labels is extracted from the multi-source time series segmented data and summarized to form a multi-source time series training sample set with health and environmental labels.

4. The device dynamic health fusion assessment method based on multi-source time-series data according to claim 3, characterized in that, Based on the operational status records, each time window segment is divided into a healthy state, a mildly degraded state, and a severely degraded state, including: Fault markers, alarm levels, and replacement component information are extracted from the operating status records, and a correspondence rule is established between the operating status records and healthy status, mild degradation status, and severe degradation status. According to the corresponding division rules, time window segments containing fault downtime records or critical component replacement records are marked as severely degraded, and the severely degraded status is written into the health label field of the corresponding time window segment. Mark time window segments that contain only alarm records or multiple minor repair records and have not experienced fault downtime as slightly degraded, and write the slightly degraded status into the health label field of the corresponding time window segment; Mark time window segments that do not contain alarm or maintenance records as healthy and write the healthy status to the health label field of the corresponding time window segment.

5. The device dynamic health fusion assessment method based on multi-source time-series data according to claim 1, characterized in that, Generating the dynamic health sequence includes: The time window degradation representations are sorted according to the timestamps, and the time window degradation representations of the same device are connected in chronological order to form a degradation feature sequence; The input tensor of the health regression network is constructed based on the degradation feature sequence, and the degradation representation of each time window and the corresponding time index are written into the input tensor. The time window with health label is degenerate and input into the health regression network via the input tensor. The regression loss function is used to constrain the deviation between the health output and the health label, and the parameters of the health regression network are updated. During the training of the health regression network, a health difference regularization term for adjacent time windows is introduced to ensure that the health output within the same degradation process remains continuous along the time direction and shows an overall downward trend. After the health regression network is trained, the degradation representation of each time window is input into the health regression network in chronological order to obtain a dynamic health sequence that corresponds one-to-one with each time window segment.

6. The device dynamic health fusion assessment method based on multi-source time-series data according to claim 5, characterized in that, The parameters of the health regression network are updated, including: Write the time window degradation representation with health label and the corresponding health label into the training batch input tensor, and maintain the one-to-one correspondence between the time window degradation representation and the health label in the training batch input tensor; The training batch is input into the tensor and then into the health regression network to obtain the health output corresponding to the degradation representation of each time window. The pairing results of the health output and the health label are recorded within the same batch. A regression loss function is constructed based on the difference between the health output and the health label corresponding to the degradation representation at each time window. Samples in mild and severe degradation states are assigned higher weights, while samples in healthy states are assigned lower weights, resulting in a weighted regression loss value. The weighted regression loss value is used as the target to backpropagate the health regression network, calculate the gradient of the network parameters of each layer, and update the health regression network parameters using a gradient descent strategy according to a preset learning rate. Repeat the above steps in multiple training rounds to iterate through all time window degradation representations with health labels until the weighted regression loss value stabilizes within the preset convergence range.

7. The device dynamic health fusion assessment method based on multi-source time-series data according to claim 5, characterized in that, The generation of the equipment degradation assessment results includes: Based on the dynamic health status sequence of multiple devices and various working conditions, the health status values ​​corresponding to the healthy state, mild degradation state, and severe degradation state are statistically analyzed to obtain the health status distribution range of each state, and a candidate set of health status thresholds is constructed. Based on the candidate set of health thresholds, the false alarm rate and false negative rate are calculated on historical samples of each device under each operating condition. A unified health threshold that achieves a trade-off between the false alarm rate and false negative rate and has a small change in the identification result under each operating condition is selected. During the online monitoring phase, real-time multi-source time-series data is collected. The real-time multi-source time-series data is processed according to timestamp alignment and time window division rules to obtain real-time time window segments. The real-time time window segments are then input into the causal representation model to output a degenerate representation of the real-time time window. The real-time time window degradation representations are concatenated in chronological order to form a real-time degradation feature sequence, and the real-time degradation feature sequence is input into the health regression network to obtain the real-time dynamic health sequence corresponding to each real-time time window segment. The health status of each time window in the real-time dynamic health status sequence is compared with the unified health status threshold, and the healthy status, mild deterioration status and severe deterioration status are determined according to the length of time that the health status is continuously below the unified health status threshold and the rate of health status decline.

8. The device dynamic health fusion assessment method based on multi-source time-series data according to claim 7, characterized in that, The health status of each time window is read from the real-time dynamic health status sequence, and the health status of each time window is compared with the unified health status threshold. A sequence of high and low health status markers relative to the unified health status threshold for each time window is recorded, wherein: Based on the high and low marker sequences, identify time periods that are continuously below the unified health threshold, count the number of time windows in each continuous time period, and convert the time window length into the length of time that is continuously below the unified health threshold. For each consecutive time period, the health status of adjacent time windows is calculated in chronological order, and the corresponding rate of decline in health status is obtained based on the health status difference and the time interval. Based on the duration of time continuously below the unified health threshold and the rate of health decline, a state division rule is constructed. Time windows with shorter durations and smaller rates of health decline are marked as mildly degraded states, while time windows with longer durations and larger rates of health decline are marked as severely degraded states. Except for the time windows marked as mildly degraded and severely degraded states, the remaining time windows are marked as healthy states.

Images