Industrial equipment state monitoring system and monitoring method based on two-way coding

Abstract

The invention provides an industrial equipment state monitoring system and method based on double-channel coding, and the system comprises a data collection and preprocessing layer which is used for collecting multi-source heterogeneous data from industrial equipment; the double-path feature coding layer is used for respectively extracting trend features and fluctuation features of the data by adopting an LSTM encoder and a CNN encoder which are parallel to each other; the feature decoding and reconstruction layer adopts a decoding structure corresponding to an encoder, an LSTM decoder reconstructs a time sequence trend through reverse expansion, a CNN decoder reconstructs local details through deconvolution operation, and two paths of reconstruction features are subjected to weighted fusion to obtain a final reconstruction sequence; and the state monitoring and diagnosis layer is used for carrying out anomaly detection and fault diagnosis based on the reconstruction error. By timely and accurately detecting the abnormal behavior of the server, the risk of fault occurrence can be reduced, the service interruption time is shortened, and the availability and reliability of the system are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent manufacturing, and in particular relates to an industrial equipment status monitoring system and a monitoring method based on dual-path coding. Background Art

[0002] With the rapid development of intelligent manufacturing, the intelligence and networking of industrial equipment are constantly increasing, and equipment operating parameters are becoming high-dimensional, multi-source, and real-time. To ensure reliable equipment operation, maintenance personnel need to monitor various equipment performance indicators in real time to promptly detect and diagnose abnormal conditions. Currently, various technical solutions exist for industrial equipment condition monitoring and fault diagnosis, but they still face many limitations in practical application. Statistical analysis-based methods, such as principal component analysis (PCA) and autoregressive models (AR), identify abnormal patterns by establishing statistical models. However, these methods make strong assumptions about data distribution and have difficulty handling non-stationary data. In real industrial environments, equipment operating conditions are often affected by factors such as changing operating conditions and environmental interference, resulting in dynamic changes in data distribution. This makes it difficult for methods based on statistical assumptions to maintain stable detection results.

[0003] Machine learning methods, such as support vector machines (SVMs) and isolation forests (IFs), detect abnormal patterns by learning from historical data. These methods require a large number of labeled abnormal samples for training. However, in real industrial scenarios, abnormal samples are often scarce and expensive to obtain, severely limiting the effectiveness of model training.

[0004] Deep learning-based methods, such as long short-term memory (LSTM) networks and generative adversarial networks (GANs), identify anomalies by learning the temporal characteristics of data. While these methods have achieved some success in specific scenarios, they still face challenges such as insufficient feature extraction and poor model interpretability.

[0005] Inadequate modeling of multi-metric correlations in equipment. The operating status of industrial equipment is determined by multiple key indicators, such as temperature, vibration, and current. These indicators are interconnected by complex physical and causal relationships. Existing methods focus on anomalies in a single or a few indicators and lack systematic modeling of the coupling relationships between these multiple indicators, resulting in high false positive rates and severe underreporting.

[0006] Poor adaptability to dynamic equipment characteristics. Industrial equipment exhibits varying dynamic characteristics under different operating conditions, and the distribution of indicators under normal conditions can also change. Existing methods mostly use static thresholds or fixed models, making it difficult to adaptively track dynamic changes in equipment status and prone to generating false alarms.

[0007] Insufficient explanation of anomalies. Beyond detecting anomalies, timely locating their causes, assessing failure risks, and recommending maintenance plans are crucial for equipment maintenance decisions. Existing methods often focus solely on anomaly detection, lacking explanation and analysis of the anomaly's mechanisms, making it difficult to provide effective decision support for maintenance personnel. Summary of the Invention

[0008] The purpose of the present invention is to provide an industrial equipment status monitoring system and monitoring method based on dual-path coding, aiming to solve the above problems.

[0009] The present invention is implemented as follows: an industrial equipment status monitoring system based on dual-path coding, comprising:

[0010] Data acquisition and preprocessing layer, used to collect multi-source heterogeneous data from industrial equipment;

[0011] The dual-path feature encoding layer uses parallel LSTM encoders and CNN encoders to extract trend features and fluctuation features of the data respectively;

[0012] The feature decoding and reconstruction layer uses a decoding structure corresponding to the encoder. The LSTM decoder reconstructs the temporal trend through reverse expansion, and the CNN decoder uses deconvolution operations to reconstruct local details. The two-way reconstructed features are weightedly fused to obtain the final reconstructed sequence.

[0013] The condition monitoring and diagnosis layer performs anomaly detection and fault diagnosis based on reconstruction errors. When an anomaly is detected, the fault type is identified by analyzing the time-frequency characteristics of the reconstruction error and combining it with the preset fault feature library. It then generates warning information including the degree of anomaly, fault type, and maintenance recommendations.

[0014] A further technical solution of the present invention is: the data acquisition and preprocessing layer collects sensor data including but not limited to temperature, vibration, current and pressure, the sampling frequency is configured according to actual needs, the collected raw data is first detected and removed for outliers, and moving median filtering is used to eliminate high-frequency noise, and then the data of different dimensions are uniformly mapped to the [0, 1] interval through the min-max normalization method, and finally a training sample sequence is constructed according to the set time window length.

[0015] A further technical solution of the present invention is: the LSTM encoder includes three layers of LSTM units, and the hidden layer dimensions are 256, 128, and 64 respectively; the CNN encoder adopts a four-layer convolution structure, the convolution kernel size is 3×3, and the number of channels is 32, 64, 128, and 256 respectively.

[0016] A further technical solution of the present invention is that the features generated by the LSTM decoder and the CNN decoder are weightedly fused to obtain the final reconstructed sequence. The system defines a reconstruction loss function based on mean square error to evaluate the reconstruction quality. The formula is: Where X and X' represent the original sequence and the reconstructed sequence respectively; the trend feature encoding path is based on LSTM and adopts an encoder-decoder structure. The LSTM encoder converts the input sequence X=x1, x2, ..., x T Mapped to a hidden state sequence H = h1, h2, ..., h T , each LSTM unit contains three gating mechanisms: input gate, forget gate and output gate. Its mathematical formula is as follows:

[0017] i t =σ(W i x t +U i h t-1 +b i )

[0018] f t =σ(W f x t +U f h t-1 +b f )

[0019] o t =σ(W o x t +U o h t-1 +b o )

[0020] Where W and U are weight matrices, b i 、b f 、b o is the bias term, σ represents the sigmoid activation function, i t 、f t 、o t They refer to the input gate, forget gate, and output gate respectively.

[0021] Another object of the present invention is to provide a monitoring method for an industrial equipment status monitoring system based on dual-path coding, comprising the following steps:

[0022] Step S1: The dual-path feature encoding layer models the long-term trend and short-term fluctuation characteristics of industrial equipment operation data respectively. It constructs a forward diffusion process through iterative noise addition, gradually destroying the spatiotemporal structure of the time series. It then trains a backward diffusion model through step-by-step denoising to learn and recover the conditional distribution of the original sample from the noise sample.

[0023] Step S2: Condition constraint diffusion process, in which the equipment operating condition information collected from the data acquisition and preprocessing layer is introduced as a condition constraint;

[0024] Step S3: Group residual vector quantization, using two independent residual vector quantizers to encode trend and fluctuation features respectively;

[0025] Step S4: Introduce the loss function gradient balancing mechanism to prevent the model from falling into local optimality.

[0026] A further technical solution of the present invention is: in step S1, given a set of normal time series samples in represents the i-th sample, which contains K channels and L time steps. N is the total number of samples in training. The training process is divided into a forward diffusion process and a reverse denoising process. In the forward diffusion process, Gaussian noise is iteratively added to the original sample x0 to gradually destroy its spatiotemporal structure and generate a series of noise samples x1, x2, ..., x T , this process can be formally expressed as:

[0027]

[0028] Among them, q(x t |x t-1 ) indicates that the noise level β t Next, from x t-1 Generate x t The conditional distribution of , that is:

[0029]

[0030] Noise level β t ∈(0,1) is a predefined hyperparameter sequence that satisfies 0<β1<β2<…<β T <1, I is the identity matrix, and the Markov property of the forward diffusion process is used to directly sample from the original sample x0:

[0031]

[0032] Among them, α t =1-β t , α t is the proportion of the signal retained in step t, is the cumulative retention rate from step 0 to step t, α s is a one-step factor in the product;

[0033] In the reverse denoising process, a diffusion model p is trained θ (x t-1 |xt ) to reverse the noise addition process, that is, to learn from the noise sample x t Recover the sample x of the previous moment t-1 , the entire reverse denoising process is expressed in the form of a Markov chain:

[0034]

[0035] Among them, p(x T ) is the prior distribution of the noise sample, which is assumed to be a standard Gaussian distribution p θ (x t-1 |x t ) is the conditional denoising distribution to be learned, with x t and time step t as input, predict x t-1 The distribution parameters of , parameterized as:

[0036]

[0037] Among them, μ θ (x t , t) and σ θ (x t , t) represent the predicted mean and standard deviation respectively, which are composed of a neural network ∈ θ (x t , t) output, according to the theory of diffusion model, the optimal denoising distribution satisfies:

[0038]

[0039] The neural network ∈ θ (x t , t) is defined as minimizing the following loss function:

[0040]

[0041] That is, let the neural network ∈ θ (x t ,t) to fit the noise residual∈, represents the average of all possible t and noise samples ε.

[0042] A further technical solution of the present invention is that in step S3, the grouped residual vector quantization includes two independent residual vector quantizers, a trend feature quantizer for encoding long-term trend information of device operation and a fluctuation feature quantizer for encoding short-term fluctuation information of device operation:

[0043] The encoding process of each group of quantizers can be expressed as: where z e is the latent feature output by the encoder, ek is the codebook vector, is the RVQ codebook, z q To quantify the results;

[0044] A structured state space model is introduced into the diffusion model. The backbone network consists of multiple S4Layers, each of which contains:

[0045] A bidirectional causal convolutional layer for extracting local spatiotemporal features;

[0046] A structured state space layer for modeling long-range dependencies. The structured state space layer is based on a linear state space model and describes the long-range relationship between the input sequence u(t) and the output sequence y(t):

[0047] x′(t)=Ax(t)+Bu(t)\y(t)=Cx(t)+Du(t)

[0048] in, is the hidden state vector, A, B, C, D are the state transfer matrices;

[0049] A feed-forward layer for nonlinear transformation and feature mixing;

[0050] A skip connection layer that directly adds the feature maps of the lower and higher layers;

[0051] By cascading multiple S4Layers, the backbone network extracts rich multi-scale spatiotemporal features from noise samples, providing better priors for the subsequent denoising process. The forward propagation of the backbone network can be expressed as:

[0052] z0=Linear(x t )\z l =S4Layerl(zl-1),e=1,2,…,L

[0053]

[0054] L is the number of layers of S4Layer, z0 is the initial hidden vector after linear projection, z l It is the output feature of the lth layer S4Layer, l is the layer number, that is, the network has L layers.

[0055] A further technical solution of the present invention is: in step S4, the overall loss function includes three parts:

[0056]

[0057] Where E(x, x*) is the overall time series reconstruction loss, E(xtrend, x*trend) is the trend reconstruction loss, E(xfluct, x*fluct) is the fluctuation reconstruction loss, λ1 is the weight of the overall sequence reconstruction term, which is used to ensure the consistency of the restored time series in the global shape, λ2 is the weight of the trend reconstruction term, emphasizing the accurate fitting of long-term, low-frequency trend signals, λ3 is the weight of the fluctuation reconstruction term, focusing on capturing short-term, high-frequency local fluctuations, λ4 is the weight of the codebook commitment loss, and Closs is the codebook commitment loss. In order to prevent the model from falling into local optimality, a gradient balancing mechanism is introduced:

[0058]

[0059] where g i is the original gradient, R is the reference norm coefficient, α is the short-term moving average coefficient, is the gradient used for back propagation after recalibration.

[0060] A further technical solution of the present invention is: in practical applications, time series data often comes from multiple sensors or devices. In order to enable the model to adaptively model the data distribution of different server individuals, entity ID embedding is introduced into the conditional input of the model, and a learnable embedding vector is maintained for each entity. Where i is the ID of the entity, d is the embedding dimension, and in each layer of the diffusion model, the entity ID is embedded into e i It is concatenated with the time step embedding Embedding(t) as an additional conditional input. The formal definition of the conditional input is:

[0061]

[0062] Accordingly, the forward propagation formula of the backbone network is modified to

[0063]

[0064] Among them, each S4Layer is z l-1 and As input, the output feature map z l .

[0065] A further technical solution of the present invention is: after obtaining the reconstructed sample set, by comparing the original sample x t The difference between the reconstructed sample and the original image at each time step is used to calculate the anomaly score, using an anomaly score based on the smoothed L1 distance:

[0066]

[0067] Among them, x t (l) represents the sample xt The value at time step l, represents the value of the s-th reconstructed sample at time step l, w(l) is a smooth window function used to control the contribution of local differences to the anomaly score, and smoothL1(·) is the smooth L1 distance function, defined as:

[0068]

[0069] Compared with the standard L1 distance, the smoothed L1 distance is smoother near the zero point, and the sample x is judged according to the anomaly score. t Whether it is an anomaly, set a threshold τ and mark the samples whose anomaly score exceeds the threshold as an anomaly:

[0070] f(x t , i)={1,ifs(x t ,i)>τ\0,otherwise

[0071] Among them, f(x t , i)=1 means sample x t is abnormal, f(x t , i)=0 means sample x t It is normal.

[0072] The beneficial effects of this invention are: by timely and accurately detecting abnormal server behavior, the risk of failures can be reduced, service interruptions can be shortened, and system availability and reliability can be improved. Anomaly detection results can provide key insights for operational tasks such as fault diagnosis, performance bottleneck analysis, and resource scheduling optimization. This helps engineers quickly locate and resolve problems, develop effective optimization strategies, maximize the effectiveness of IT infrastructure, and reduce operating costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 It is the main flow chart of the method of the present invention;

[0074] Figure 2 It is the overall architecture diagram of the system of the present invention;

[0075] Figure 3 Schematic diagram of the encoder and decoder structure of the present invention;

[0076] Figure 4 This is a diagram of the grouped residual vector quantization structure of the present invention;

[0077] Figure 5 is a schematic diagram of a cascaded structured state space model of the present invention;

[0078] Figure 6 This is the experimental data comparison curve of the present invention Figure 1 ;

[0079] Figure 7 This is the experimental data comparison curve of the present invention Figure 2 . DETAILED DESCRIPTION

[0080] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0081] This paper proposes an industrial equipment status monitoring system and method based on dual-path encoding. This method innovatively combines the advantages of long short-term memory (LSTM) networks and convolutional neural networks (CNNs), constructing a dual-path encoding structure to extract trend and fluctuation characteristics of equipment operating data. This method achieves efficient feature compression through residual vector quantization (RVQ). This method can accurately capture the state changes of industrial equipment at different time scales, providing reliable technical support for equipment anomaly detection and fault diagnosis.

[0082] The industrial equipment condition monitoring and fault diagnosis system proposed in this paper adopts a hierarchical architecture. From bottom to top, the system comprises a data acquisition and preprocessing layer, a dual-path feature encoding layer, a feature decoding and reconstruction layer, and a condition monitoring and diagnosis layer. Standardized interfaces enable data exchange and information transfer between these layers. The data acquisition and preprocessing layer is responsible for collecting multi-source heterogeneous data from industrial equipment. The system supports the acquisition of various sensor data types, such as temperature, vibration, current, and pressure. The sampling frequency can be configured according to actual needs, typically between 100 Hz and 1 kHz. The collected raw data is first detected and removed for outliers, and a moving median filter is used to eliminate high-frequency noise. Then, a min-max normalization method is used to uniformly map the data of different dimensions to the [0, 1] interval for ease of subsequent processing. Finally, a training sample sequence is constructed based on a set time window length (typically 1024 sampling points).

[0083] The dual-path feature encoding layer is the core component of the system. It uses a parallel LSTM encoder and CNN encoder to extract data trend and fluctuation characteristics, respectively. The LSTM encoder consists of three layers of LSTM units, with hidden layer dimensions of 256, 128, and 64, respectively, to capture long-term temporal dependencies in the data. The CNN encoder uses a four-layer convolutional structure with a 3×3 kernel size and 32, 64, 128, and 256 channels, respectively, to extract local fluctuation patterns. The dual-path features are compressed by the RVQ module to obtain a low-dimensional representation, achieving a compression rate of over 90%.

[0084] The feature decoding and reconstruction layer uses a decoding structure corresponding to the encoder. The LSTM decoder reconstructs the temporal trend by reverse expansion, while the CNN decoder reconstructs local details using deconvolution operations. The two-way reconstructed features are weighted and fused to obtain the final reconstructed sequence. The system defines a reconstruction loss function based on mean square error to evaluate the reconstruction quality, as shown in formula (1):

[0085]

[0086] where X and X' represent the original sequence and the reconstructed sequence, respectively.

[0087] The condition monitoring and diagnosis layer uses reconstruction errors to detect anomalies and diagnose faults. The system employs an adaptive threshold mechanism to dynamically adjust alarm thresholds based on the statistical characteristics of historical data. When an anomaly is detected, the system analyzes the time-frequency characteristics of the reconstruction error and, in combination with a pre-defined fault signature library, identifies the fault type. It then generates a warning message containing the severity of the anomaly, the fault type, and maintenance recommendations.

[0088] The detailed structure of the dual-pass coding architecture is as follows Figure 2 As shown in Figure 2, it mainly consists of three parts: a trend feature encoding path, a fluctuation feature encoding path, and a feature fusion mechanism. This innovative dual-path structure design fully considers the multi-scale characteristics of industrial equipment operation data and can simultaneously capture long-term evolution trends and instantaneous fluctuation characteristics.

[0089] The trend feature encoding path is based on LSTM and adopts an encoder-decoder structure. The encoder takes the input sequence X=x1, x2, ..., x T Mapped to a hidden state sequence H = h1, h2, ..., h T Each LSTM unit contains three gating mechanisms: input gate, forget gate, and output gate. Its mathematical description is shown in formulas (2)-(4):

[0090] i t =σ(W i x t +U i h t-1 +b i )(2)

[0091] f t =σ(W f x t +U f h t-1 +b f )(3)

[0092] o t =σ(W o x t +U o h t-1+b o )(4)

[0093] Where W and U are weight matrices, b is the bias term, σ represents the sigmoid activation function, i t 、f t 、o t These gates represent the input gate, forget gate, and output gate, respectively. Through this gating mechanism, LSTM can effectively model long-term dependencies and capture the overall evolution trend of data.

[0094] In encoder and decoder design, Figure 3 As shown, the encoder and decoder employ a symmetric convolutional neural network architecture. The trend encoder / decoder (Encoder LSTM / Decoder LSTM) incorporates residual convolutional units with an expandable dilation range, a two-layer LSTM for sequence modeling, and strided convolution for compression, achieving varying compression rates. Furthermore, the fluctuation encoder / decoder (Encoder Block / Decoder Block) shares a similar architecture to the trend encoder, but lacks the LSTM layer and has a fixed dilation range of 1.

[0095] In the forward diffusion process, the original industrial equipment time series data x0 is converted into pure noise x by gradually adding Gaussian noise. t The process can be expressed as:

[0096]

[0097] Among them, the conditional distribution of each step is:

[0098]

[0099] After training is completed, the model gradually recovers the normal operating data sequence of industrial equipment through the reverse denoising process.

[0100] In order to better capture the long-range dependencies of industrial equipment data, the present invention introduces a structured state space model (S4) into the diffusion model. Its basic form is:

[0101]

[0102] y(t)=Ch(t)+Du(t)

[0103] Where h(t) is the hidden state vector, u(t) is the input vector, y(t) is the output vector, and A, B, C, and D are learnable parameter matrices.

[0104] like Figure 4 As shown in Figure 2, grouped residual vector quantization (GRVQ) consists of two independent groups of residual vector quantizers:

[0105] Trend feature quantizer: used to encode long-term trend information of equipment operation;

[0106] Fluctuation characteristic quantizer: used to encode short-term fluctuation information of equipment operation;

[0107] The encoding process of each group of quantizers can be expressed as: where z e is the latent feature output by the encoder, e k is the codebook vector, is the RVQ codebook, z q To quantify the results.

[0108] Loss function design:

[0109] The overall loss function consists of three parts:

[0110]

[0111] Where E(x, x*) is the overall time series reconstruction loss, E(xtrend, x*trend) is the trend reconstruction loss, E(xfluct, x*fluct) is the fluctuation reconstruction loss, λ1 is the weight of the overall sequence reconstruction term, which is used to ensure the consistency of the restored time series in the global shape, λ2 is the weight of the trend reconstruction term, which emphasizes the accurate fitting of long-period, low-frequency trend signals, λ3 is the weight of the fluctuation reconstruction term, which focuses on capturing short-period, high-frequency local fluctuations, λ4 is the weight of the codebook commitment loss, which is used to constrain the encoder output to be consistent with the discrete codebook during vector quantization, thereby ensuring the stability of model training, and Closs is the codebook commitment loss.

[0112] To prevent the model from falling into local optimality, a gradient balancing mechanism is introduced:

[0113]

[0114] where g i is the original gradient, R is the reference norm coefficient, α is the short-term moving average coefficient (0.99), is the gradient used for back propagation after recalibration.

[0115] The goal of the training phase is to learn the normal multi-indicator time series distribution from massive server machine data. Given a batch of normal time series samples. The goal is to learn the distribution of normal time series data. Given a set of normal time series samples in represents the i-th sample, contains K channels, L time steps, and N is the total number of samples in training. The training process is divided into two sub-stages: forward diffusion and backward denoising.

[0116] In the forward diffusion process, Gaussian noise is iteratively added to the original sample x0, gradually destroying its spatiotemporal structure to generate a series of noise samples x1, x2, ..., x T This process can be formally expressed as:

[0117]

[0118] Among them, q(x t |x t-1 ) indicates that the noise level β t Next, from x t-1 Generate x t The conditional distribution of , that is:

[0119]

[0120] Here, the noise level β t ∈(0,1) is a predefined hyperparameter sequence that satisfies 0<β1<β2<…<β T < 1. I is the identity matrix.

[0121] In order to more efficiently calculate the noise sample x at any time step t t , we can use the Markov property of the forward diffusion process to directly sample from the original sample x0:

[0122] Among them, α t =1-β t , α t is the proportion of the signal retained in step t, is the cumulative retention rate from step 0 to step t, α s is a one-step factor in the product.

[0123] In the reverse denoising process, a diffusion model p is trained θ (x t-1 |x t ) to reverse the noise addition process, that is, to learn from the noise sample x t Recover the sample x of the previous moment t-1 The entire reverse denoising process can be expressed in the form of a Markov chain:

[0124]

[0125] Among them, p(x T ) is the prior distribution of the noise sample, usually assumed to be a standard Gaussian distribution .p θ (x t-1 |x t) is the conditional denoising distribution to be learned, which is x t and time step t as input, predict x t-1 In the present invention, it is parameterized as:

[0126]

[0127] Among them, μ θ (x t , t) and σ θ (x t , t) represent the predicted mean and standard deviation respectively, which are composed of a neural network ∈ θ (x t , t) output. According to the theory of diffusion model, the optimal denoising distribution satisfies:

[0128]

[0129] Therefore, the neural network ∈ θ (x t , t) is defined as minimizing the following loss function:

[0130]

[0131] That is, let the neural network ∈ θ (x t ,t) to fit the noise residual∈, represents the average of all possible t and noise samples ε.

[0132] In order to better capture the long-range dependencies and complex interactions of server data, the present invention introduces a cascaded structured state space model (S4) into the backbone network of the diffusion model. Figure 5 shown.

[0133] Specifically, the backbone network consists of multiple S4Layers, each of which contains:

[0134] (1) A bidirectional causal convolution layer to extract local spatiotemporal features.

[0135] (2) A structured state space layer (S4) for modeling long-range dependencies. The S4 layer is based on a linear state space model and can efficiently characterize the long-range relationship between the input sequence u(t) and the output sequence y(t):

[0136] x′(t)=Ax(t)+Bu(t)\y(t)=Cx(t)+Du(t)

[0137] in, is the hidden state vector, and A, B, C, and D are the state transition matrices. By cleverly parameterizing these matrices, S4 can characterize complex long-range interactions.

[0138] Where L is the number of S4Layer layers.

[0139] (3) A feedforward layer for nonlinear transformation and feature mixing.

[0140] (4) A skip connection directly adds the feature maps of the low and high layers, which can prevent the gradient from disappearing and fuse the spatiotemporal representations of different scales.

[0141] By cascading multiple S4Layers, the backbone network can extract rich multi-scale spatiotemporal features from noise samples, providing better priors for the subsequent denoising process. The forward propagation of the backbone network can be expressed as:

[0142] z0=Linear(x t )\z l =S4Layerl(zl-1),l=1,2,…,L

[0143]

[0144] L is the number of layers of S4Layer, z0 is the initial hidden vector after linear projection, z l is the output feature of the eth layer S4 layer, and l is the layer number, meaning the network has L layers. In practical applications, time series data often comes from multiple sensors or devices, which may have different acquisition modes and abnormal behaviors. To enable the model to adaptively model the data distribution of different servers, this paper introduces entity ID embedding in the model's conditional input.

[0145] Specifically, a learnable embedding vector is maintained for each entity Where i is the ID of the entity and d is the embedding dimension. In each layer of the diffusion model, the entity ID is embedded into e i It is concatenated with the time step embedding Embedding(t) as an additional conditional input. This enables the model to dynamically adjust its feature extraction and denoising strategies based on the identity of the entity, thereby better adapting to the data distribution of different entities. The formal definition of the conditional input is:

[0146]

[0147] Accordingly, the forward propagation formula of the backbone network can be modified as follows:

[0148]

[0149] Among them, each S4Layer is z l-1 and As input, the output feature map z l .

[0150] Through the above design, the diffusion model of the present invention can take into account both local and global spatiotemporal features, adaptively model the normal data distribution of multiple entities, and lay a solid foundation for downstream anomaly detection tasks. In the following process, the complete process of model training is shown.

[0151] Algorithm 1: Training process

[0152] Input: Normal time series samples Number of training rounds M, batch size B, number of time steps T, noise level sequence β t t=1 T , entity ID embedded

[0153] Output: trained model parameters θ.

[0154] process:

[0155] a. Initialize model parameters θ

[0156] b. For m = 1 to M:

[0157] i. Randomly sample a batch of samples

[0158] ii. Random sampling noise and time step t~Uniform(1,…,T)

[0159] iii. Generate noise samples

[0160] iv. Calculate losses

[0161] v. Calculate gradients through backpropagation And update the parameter θ

[0162] In the above process, each training step first randomly samples a batch of normal samples from the dataset Then randomly sample noise ∈ and time step t to generate the corresponding noise samples Next, the noise sample, time step, and entity ID are input into the model, the reconstruction loss is calculated, and the model parameters are updated through backpropagation. This process is repeated M times to obtain the trained model parameters θ.

[0163] In the testing phase, the trained diffusion model is used to detect anomalies in new time series samples. And its corresponding entity IDi, the detection process can be divided into the following steps:

[0164] First, the sample to be tested x t The trained diffusion model is input with entity IDi, and a series of reconstructed samples are generated through the reverse denoising process. Specifically, at each denoising step, from the prior distribution p(x T ) Sample a noise sample x T , and then iteratively apply the denoising distribution pθ(x t-1 |x t ), until the final reconstructed sample hat{x}_0 is obtained:

[0165]

[0166] Among them, the mean μ θ (x t , t, i) and standard deviation σ θ (x t , t,i) is determined by the output of the model ∈ θ (x t , t, i). Repeating this process several times, we can get the empirical distribution p of the reconstructed sample θ (x0|x t , i). The following is the detailed sampling process of the present invention.

[0167] Algorithm 2: Sampling process

[0168] Input: sample x to be tested t , entity IDi, time step number T, noise level sequence Model parameters θ after training, number of sampling S;

[0169] Output: Reconstructed sample set

[0170] process:

[0171] a. Initialize model parameters θ

[0172] b. For t = T to 1:

[0173] i. Calculate the mean

[0174] ii. Calculate the standard deviation

[0175] iii. Sampling from the denoised distribution σθ (x t ,t,i)I)

[0176] c. Get the sth reconstruction sample

[0177] Get the reconstructed sample set After that, we can compare the original sample x t The anomaly score is calculated by the difference between the reconstructed sample and the original sample at each time step. In order to capture local abnormal patterns, the present invention adopts an anomaly score based on the smoothed L1 distance:

[0178]

[0179] Among them, x t (l) represents the sample x t The value at time step l, represents the value of the s-th reconstructed sample at time step l. w(l) is a smooth window function used to control the contribution of local differences to the anomaly score. smoothL1(·) is the smooth L1 distance function, defined as:

[0180]

[0181] Compared with the standard L1 distance, the smoothed L1 distance is smoother near the zero point and can alleviate the impact of outliers on the distance metric. Finally, according to the anomaly score s(x t , i) judge the sample x t Is it an anomaly? A simple judgment method is to set a threshold τ and mark samples with anomaly scores exceeding the threshold as anomalies:

[0182] f(x t , i)={1,ifs(x t ,i)>τ\0,otherwise

[0183] Among them, f(x t , i)=1 means sample x t is abnormal, f(x t , i)=0 means sample x t The threshold τ can be adjusted according to the anomaly detection performance on the validation set to achieve the desired precision and recall.

[0184] Algorithm 3: Anomaly Detection Process

[0185] Input: sample x to be tested t , entity IDi, trained model parameters θ, anomaly threshold τ;

[0186] Output: abnormal label f(x t, i).

[0187] process:

[0188] a. Use Algorithm 2 to generate a set of reconstructed samples

[0189] b. Calculate anomaly score

[0190] c. If s(x t , i)>τ, then:

[0191] if(x t , i) = 1 (abnormal)

[0192] Otherwise, if s(x t ,i)≤τ, then:

[0193] ii.f(x t , i) = 0 (normal)

[0194] Through the above process, the diffusion model of the present invention can adaptively learn the normal time series distribution of different entities and effectively detect abnormal samples through reconstruction errors in the testing phase, providing a novel, universal and robust solution for the time series anomaly detection task.

[0195] exist Figure 6 and Figure 7 The figure below demonstrates the anomaly detection performance of the present invention on different datasets. The blue curve represents the original test sample, the orange curve represents the reconstructed sample, and the red curve represents the smoothed L1 distance. It can be seen that the reconstructed sample fits the normal data pattern well, while the smoothed L1 distance exhibits a significant peak within the anomaly interval. Introducing entity ID embedding further improves the reconstruction performance and achieves higher discriminability of the anomaly score.

[0196] Table 1 quantitatively compares the anomaly detection performance of our proposed method against other baseline methods on three datasets, including precision, recall, and F1-score. As can be seen, our diffusion model achieves optimal performance on all datasets, particularly after the introduction of entity ID embedding, where performance is further improved, reaching state-of-the-art performance. This fully demonstrates the effectiveness and superiority of our proposed method for time series anomaly detection.

[0197] Table 1

[0198]

[0199] We also explored the impact of the choice of window function in the smoothed L1 distance on detection performance. Table 2 lists the intersection over union (IoU) of detected anomalies and true anomalies for different window functions and window sizes. As can be seen, using a Hanning window with a window size of 25 achieves an optimal IoU of 92.4%, demonstrating that the smoothed L1 distance effectively covers the contextual anomaly range.

[0200] Table 2

[0201]

[0202] Based on the above method, a server multi-indicator time series anomaly detection system was built. This system mainly includes:

[0203] 1. Data access module: connects to the server monitoring data source, collects multi-indicator time series data in real time, and performs pre-processing such as cleaning and alignment.

[0204] 2. Feature Engineering Module: Performs feature engineering operations such as data normalization, feature selection, and time window segmentation to generate input data for the model.

[0205] 3. Anomaly detection module: Based on the offline trained conditional diffusion model, it reconstructs and samples the real-time data and calculates the anomaly score, and outputs the anomaly discrimination results.

[0206] 4. Visualization module: Displays the monitoring curves and reconstruction curves of various server indicators in real time, and highlights abnormal intervals to assist operation and maintenance personnel in quickly locating problems.

[0207] 5. Alarm module: Based on the anomaly detection results, combined with preset thresholds and business rules, it automatically generates anomaly alarms and notifies relevant personnel via email, SMS, etc.

[0208] 6. Root cause analysis module: After an alarm is generated, it automatically associates logs, call chains and other data, applies statistical and causal reasoning algorithms, and assists in locating the root cause of the anomaly and guiding its handling.

[0209] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0210] By introducing the conditional diffusion model, we adaptively learn the time series distribution of multiple indicators under normal server conditions, building a reference system that comprehensively describes server behavior. During the testing phase, we automatically identify various anomalies using reconstruction error and smoothed distance, significantly improving detection accuracy, recall, and comprehensiveness.

[0211] By introducing a structured state space model into the diffusion model backbone network, a multi-scale representation of multi-indicator time series data from servers is constructed through the cascaded S4 layer. This effectively models the long-range dependencies and complex interactions between key indicators, overcoming the inadequacy of existing methods in modeling multi-indicator correlations.

[0212] The innovative introduction of entity ID embedding enables the diffusion model to adaptively learn the behavior patterns of different server individuals. This not only improves detection accuracy but also significantly enhances the model's generalization and robustness, effectively addressing the challenges posed by varying server workloads and individual differences.

[0213] Anomaly scores are calculated based on the smoothed L1 distance, which not only improves overall anomaly detection but also makes it more sensitive to local anomalies of key indicators. This makes up for the shortcomings of existing methods in capturing isolated and weak anomalies.

[0214] By leveraging the reconstruction capabilities of the generative diffusion model and comparing the original and reconstructed curves, the anomaly occurrence period, anomaly severity, and key anomaly indicators can be intuitively displayed. This significantly improves the interpretability of the detection results and provides important clues for operations and maintenance personnel to quickly locate and diagnose the root cause.

[0215] We have implemented a complete server anomaly detection system, providing an end-to-end closed loop from data access, anomaly detection, visual display, alarm notification, and root cause analysis. This system utilizes a microservices architecture and a configurable design, offering excellent versatility, scalability, and real-time performance. It can flexibly adapt to different scenarios and effectively reduce operation and maintenance costs.

[0216] The overall solution adopts an unsupervised learning paradigm, automatically constructing abnormal references by learning massive amounts of normal data. Only a very small number of abnormal samples are required for verification and tuning, overcoming the problems of sparse abnormal samples and high labeling costs, and greatly improving the practicality of the solution.

[0217] Therefore, the present invention has significant application value in the field of anomaly detection of server machine data. By timely and accurate detection of abnormal server behavior, the risk of failure can be reduced, service interruption time can be reduced, and the availability and reliability of the system can be improved. Anomaly detection results can provide key basis for operation and maintenance tasks such as fault diagnosis, performance bottleneck analysis, and resource scheduling optimization, helping engineers quickly locate and solve problems, formulate effective optimization strategies, maximize the effectiveness of IT infrastructure, and save operating costs. The present invention provides new ideas and technical support for intelligent operation and maintenance of data centers and AIOps practices.

[0218] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An industrial equipment status monitoring system based on dual-path coding, characterized in that: include: Data acquisition and preprocessing layer, used to collect multi-source heterogeneous data from industrial equipment; The dual-path feature encoding layer uses parallel LSTM encoders and CNN encoders to extract trend features and fluctuation features of the data respectively; The feature decoding and reconstruction layer uses a decoding structure corresponding to the encoder. The LSTM decoder reconstructs the temporal trend through reverse expansion, and the CNN decoder uses deconvolution operations to reconstruct local details. The two-way reconstructed features are weightedly fused to obtain the final reconstructed sequence. The condition monitoring and diagnosis layer performs anomaly detection and fault diagnosis based on reconstruction errors. When an anomaly is detected, the fault type is identified by analyzing the time-frequency characteristics of the reconstruction error and combining it with a preset fault feature library. It then generates early warning information including the degree of anomaly, fault type, and maintenance recommendations.

2. The industrial equipment status monitoring system based on dual-path coding according to claim 1 is characterized in that: The data acquisition and preprocessing layer collects sensor data including but not limited to temperature, vibration, current and pressure. The sampling frequency is configured according to actual needs. The collected raw data is first detected and removed for outliers, and a moving median filter is used to eliminate high-frequency noise. Then, the data of different dimensions are uniformly mapped to the [0, 1] interval through the min-max normalization method. Finally, a training sample sequence is constructed according to the set time window length.

3. The industrial equipment status monitoring system based on dual-path coding according to claim 1 is characterized in that: The LSTM encoder contains three layers of LSTM units with hidden layer dimensions of 256, 128, and 64 respectively. The CNN encoder adopts a four-layer convolution structure with a convolution kernel size of 3×3 and the number of channels of 32, 64, 128, and 256 respectively.

4. The industrial equipment status monitoring system based on dual-path coding according to claim 1 is characterized in that: The features generated by the LSTM decoder and the CNN decoder are weighted and fused to obtain the final reconstructed sequence. The system defines a reconstruction loss function based on mean square error to evaluate the reconstruction quality. The formula is: Where X and X' represent the original sequence and the reconstructed sequence respectively; The trend feature encoding path is based on LSTM and adopts an encoder-decoder structure. The LSTM encoder converts the input sequence X=x1, x2, ..., x T Mapped to a hidden state sequence H = h1, h2, ..., h T , each LSTM unit contains three gating mechanisms: input gate, forget gate and output gate. Its mathematical formula is as follows: i t =σ(W i x t +U i h t-1 +b i ) f t =σ(W f x t +U f h t-1 +b f ) the t =σ(W o x t +U o h t-1 +b o ) Where W and U are weight matrices, b i 、b f 、b o is the bias term, σ represents the sigmoid activation function, i t 、f t 、o t They refer to the input gate, forget gate, and output gate respectively.

5. A monitoring method for an industrial equipment status monitoring system based on dual-path coding according to any one of claims 1 to 4, characterized in that: The following steps are involved: Step S1: The dual-path feature encoding layer models the long-term trend and short-term fluctuation characteristics of industrial equipment operation data respectively. It constructs a forward diffusion process through iterative noise addition, gradually destroying the spatiotemporal structure of the time series. It then trains a backward diffusion model through step-by-step denoising to learn and recover the conditional distribution of the original sample from the noise sample. Step S2: Condition constraint diffusion process, in which the equipment operating condition information collected from the data acquisition and preprocessing layer is introduced as a condition constraint; Step S3: Group residual vector quantization, using two independent residual vector quantizers to encode trend and fluctuation features respectively; Step S4: Introduce the loss function gradient balancing mechanism to prevent the model from falling into local optimality.

6. The monitoring method of the industrial equipment status monitoring system based on dual-path coding according to claim 5, characterized in that: In step S1, a set of normal time series samples is given in represents the i-th sample, which contains K channels and L time steps. N is the total number of samples in training. The training process is divided into a forward diffusion process and a reverse denoising process. In the forward diffusion process, Gaussian noise is iteratively added to the original sample x0 to gradually destroy its spatiotemporal structure and generate a series of noise samples x1, x2, ..., x T , this process can be formally expressed as: Among them, q(x t |x t-1 ) indicates that the noise level β t Next, from x t-1 Generate x t The conditional distribution of , that is: Noise level β t ∈(0,1) is a predefined hyperparameter sequence that satisfies 0<β1<β2<…<β T <1, I is the identity matrix, and the Markov property of the forward diffusion process is used to directly sample from the original sample x0: Among them, α t =1-β t , α t is the proportion of the signal retained in step t, is the cumulative retention rate from step 0 to step t, α s is a one-step factor in the product; In the reverse denoising process, a diffusion model p is trained θ (x t-1 |x t ) to reverse the noise addition process, that is, to learn from the noise sample x t Recover the sample x of the previous moment t-1 , the entire reverse denoising process is expressed in the form of a Markov chain: Among them, p(x T ) is the prior distribution of the noise sample, which is assumed to be a standard Gaussian distribution p θ (x t-1 |x t ) is the conditional denoising distribution to be learned, with x t and time step t as input, predict x t-1 The distribution parameters of , parameterized as: Among them, μ θ (x t , t) and σ θ (x t , t) represent the predicted mean and standard deviation respectively, which are composed of a neural network ∈ θ (x t , t) output, according to the theory of diffusion model, the optimal denoising distribution satisfies: The neural network ∈ θ (x t , t) is defined as minimizing the following loss function: That is, let the neural network ∈ θ (x t ,t) to fit the noise residual∈, represents the average of all possible t and noise samples ε.

7. The monitoring method of the industrial equipment status monitoring system based on dual-path coding according to claim 5, characterized in that: In step S3, the grouped residual vector quantization includes two independent groups of residual vector quantizers, a trend feature quantizer for encoding long-term trend information of device operation and a fluctuation feature quantizer for encoding short-term fluctuation information of device operation: The encoding process of each group of quantizers can be expressed as: where z e is the latent feature output by the encoder, e k is the codebook vector, is the RVQ codebook, z q To quantify the results; A structured state space model is introduced into the diffusion model. The backbone network consists of multiple S4Layers, each of which contains: A bidirectional causal convolutional layer for extracting local spatiotemporal features; A structured state space layer for modeling long-range dependencies. The structured state space layer is based on a linear state space model and describes the long-range relationship between the input sequence u(t) and the output sequence y(t): x′(t)=Ax(t)+Bu(t)\y(t)=Cx(t)+Du(t) in, is the hidden state vector, A, B, C, D are the state transfer matrices; A feed-forward layer for nonlinear transformation and feature mixing; A skip connection layer that directly adds the feature maps of the lower and higher layers; By cascading multiple S4Layers, the backbone network extracts rich multi-scale spatiotemporal features from noise samples, providing better priors for the subsequent denoising process. The forward propagation of the backbone network can be expressed as: L is the number of layers of S4Layer, z0 is the initial hidden vector after linear projection, It is Layer S4Layer outputs features, It is the layer number, that is, there are L layers in the network.

8. The monitoring method of the industrial equipment status monitoring system based on dual-path coding according to claim 5, characterized in that: In step S4, the overall loss function consists of three parts: Where E(x, x*) is the overall time series reconstruction loss, E(xtrend, x*trend) is the trend reconstruction loss, E(xfluct, x*fluct) is the fluctuation reconstruction loss, λ1 is the weight of the overall sequence reconstruction term, which is used to ensure the consistency of the restored time series in the global shape, λ2 is the weight of the trend reconstruction term, emphasizing the accurate fitting of long-term, low-frequency trend signals, λ3 is the weight of the fluctuation reconstruction term, focusing on capturing short-term, high-frequency local fluctuations, λ4 is the weight of the codebook commitment loss, and Closs is the codebook commitment loss. In order to prevent the model from falling into local optimality, a gradient balancing mechanism is introduced: where g i is the original gradient, R is the reference norm coefficient, α is the short-term moving average coefficient, is the gradient used for back propagation after recalibration.

9. The monitoring method of the industrial equipment status monitoring system based on dual-path coding according to claim 5, characterized in that: In practical applications, time series data often comes from multiple sensors or devices. In order to enable the model to adaptively model the data distribution of different server individuals, entity ID embedding is introduced in the conditional input of the model, maintaining a learnable embedding vector for each entity. Where i is the ID of the entity, d is the embedding dimension, and in each layer of the diffusion model, the entity ID is embedded into e i It is concatenated with the time step embedding Embedding(t) as an additional conditional input. The formal definition of the conditional input is: Accordingly, the forward propagation formula of the backbone network is modified to Among them, each S4Layer is and As input, output feature map 10. The monitoring method of the industrial equipment status monitoring system based on dual-path coding according to claim 5, characterized in that: After obtaining the reconstructed sample set, by comparing the original sample x t The difference between the reconstructed sample and the original image at each time step is used to calculate the anomaly score, using an anomaly score based on the smoothed L1 distance: Among them, x t (l) represents the sample x t The value at time step l, represents the value of the s-th reconstructed sample at time step l, w(l) is a smooth window function used to control the contribution of local differences to the anomaly score, and smoothL1(·) is the smooth L1 distance function, defined as: Compared with the standard L1 distance, the smoothed L1 distance is smoother near the zero point, and the sample x is judged according to the anomaly score. t Whether it is an anomaly, set a threshold τ and mark the samples whose anomaly score exceeds the threshold as an anomaly: f(x t ,i)={1,ifs(x t ,i)>τ\0,otherwise Among them, f(x t , i)=1 means sample x t is abnormal, f(x t , i)=0 means sample x t It is normal.

Images