An industrial mobile robot periodic feature signal anomaly detection method

Abstract

The application discloses an abnormality detection method for periodic characteristic signals of an industrial mobile robot, which comprises the following steps: firstly, obtaining original time series data of the mobile robot from an industrial database and performing data preprocessing; then, constructing an LSTM-AE framework model, selecting the preprocessed time series data under the stable operation of the industrial mobile robot to reconstruct and train the model, optimizing parameters by using a back propagation algorithm until the evaluation index approaches the optimal value; then, calculating the error E of the reconstructed sequence, fitting the error Gaussian distribution, and calculating the Gaussian confidence interval GCI; finally, feeding real-time data into the model for reconstruction, comparing the reconstruction error with the GCI, and realizing the abnormality detection of the industrial mobile robot. The application realizes real-time abnormality detection and early warning of the periodic characteristic signals of the industrial mobile robot, and has the advantages of high detection precision and work efficiency and low maintenance cost.

Description

Technical Field

[0001] This invention relates to the field of state monitoring and fault diagnosis of periodic characteristic signals for complex industrial mobile robots, and in particular to a method for detecting abnormal periodic characteristic signals of industrial mobile robots. Background Technology

[0002] With the rapid development of the Industrial Internet of Things (IIoT), industrial mobile robots face increasingly complex management, maintenance, and operation, as well as rising application costs. Simultaneously, metal fatigue caused by prolonged, uninterrupted operation in their working environment leads to a higher probability of equipment failure. In recent years, the application of artificial intelligence and big data technologies in data mining has enabled the extraction of equipment status information from monitoring data during production processes. This allows for pre-failure maintenance of various industrial production equipment, providing new opportunities for studying the dynamic behavior and safe operation of complex systems.

[0003] Applying fault diagnosis technology to equipment monitoring and diagnosis can quickly detect machine faults, thereby preventing serious equipment accidents, reducing casualties, minimizing environmental pollution, eliminating potential hazards, and avoiding losses. Therefore, fault diagnosis technology plays a vital role in the upgrading and transformation of the manufacturing industry and has a significant driving effect on national economic development.

[0004] Long Short-Term Memory (LSTM) networks encode and store data through memory units when processing time-series data, and solve the gradient vanishing and gradient exploding problems in the training process of long-term sequences, enabling efficient extraction of features from periodic signals.

[0005] An autoencoder (AE) is a type of artificial neural network used in semi-supervised and unsupervised learning to reconstruct input signals. An AE consists of an encoder and a decoder. The encoder extracts features from periodic signals, reducing data dimensionality, while the decoder decodes the features output by the encoder into the same form as the input, obtaining the reconstructed periodic signal. Summary of the Invention

[0006] The purpose of this invention is to provide a method for detecting anomalies in periodic characteristic signals of industrial mobile robots that can achieve real-time anomaly detection and early warning of periodic characteristic signals, with high detection accuracy and work efficiency, and low maintenance cost.

[0007] The technical solution to achieve the objective of this invention is: a method for detecting abnormal periodic characteristic signals of industrial mobile robots, comprising the following steps:

[0008] Step 1: Obtain periodic time series data of industrial mobile robots and perform data preprocessing to obtain training dataset, validation dataset and test dataset for the model;

[0009] Step 2: Utilize the ability of the LSTM deep model to extract time series characteristics, and construct an LSTM-AE anomaly detection model based on the autoencoder (AE) reconstruction principle.

[0010] Step 3: Use the preprocessed training dataset as input to the LSTM-AE anomaly detection model for reconstruction training. Optimize the parameters using the backpropagation algorithm until the evaluation metric is optimal. Save the LSTM-AE anomaly detection model and its parameters.

[0011] Step 4: Using the trained LSTM-AE anomaly detection model, reconstruct the periodic feature signal of the mobile robot under normal working conditions, calculate the reconstruction error, and calculate the GCI of the periodic feature signal based on the reconstruction error.

[0012] Step 5: Input the real-time acquired periodic feature signal into the trained LSTM-AE anomaly detection model. By comparing the reconstruction error output by the LSTM-AE anomaly detection model with the GCI, determine whether the periodic feature signal is abnormal.

[0013] Further, step 1 involves acquiring periodic time series data of the industrial mobile robot and performing data preprocessing to obtain the model's training dataset, validation dataset, and test dataset, as detailed below:

[0014] Step 1.1: Select periodic time series data of industrial mobile robots under suitable working conditions, and extract the required data using the equal time interval method;

[0015] Step 1.2: Based on the periodicity of the periodic characteristic signal, the collected time series is divided into samples of fixed length, with each sample serving as a time unit and L being the sample length.

[0016] Step 1.3: Set N to be the sample length of the extracted state periodic time series, then the periodic time series X = [X1, X2, X3, ..., X...]. N The format after dividing by sample length is as follows:

[0017]

[0018] Step 1.4: Standardize the data for each time unit using the MinMaxScaler() function, then convert it to tensor form to obtain the model's training dataset Tr_D, validation dataset Va_D, and test dataset Te_D. The sample length of the training dataset Tr_D is N1, the sample length of the validation dataset Va_D is N2, and the sample length of the test dataset Te_D is N3, in the following format:

[0019]

[0020] Where tensor(·) represents a tensor. To represent a sample of data, Represents the training dataset. This represents the validation dataset. Let N represent the test dataset, and let i and j represent the starting points of the validation dataset and the test dataset, respectively. The datasets are mutually exclusive, and N1 + N2 + N3 = N.

[0021] Furthermore, step 2 utilizes the ability of the LSTM deep model to extract time series characteristics and constructs an LSTM-AE anomaly detection model based on the autoencoder (AE) reconstruction principle, as detailed below:

[0022] Using LSTM units as neurons in the encoding and decoding process, and utilizing the control gates and memory gates in the LSTM units, and based on the autoencoder (AE) reconstruction principle, an LSTM-AE anomaly detection model is constructed to learn the characteristics of time-periodic sequences.

[0023] Furthermore, the LSTM deep model described in step 2 is as follows:

[0024] LSTM deep model is a type of RNN network that introduces the concept of gates to update state variables. It is controlled by three gates: the input gate, the forget gate, and the output gate. The input gate controls the network's input, the forget gate controls the memory units, giving LSTM long-term memory capabilities, and the output gate controls the network's output. The calculation formula for an LSTM unit is as follows:

[0025] i t =σ(W i *[h t-1 ,x t ]+b i )

[0026] f t =σ(W f *[h t-1 ,x t ]+b f )

[0027] ot =σ(W o *[h t-1 ,x t ]+b o )

[0028] C t =tanh(W g *[h t-1 ,x t ]+b g )

[0029] C t+1 =σ(f t *C t-1 +i t *C t )

[0030] h t =tanh(C t+1 )*o t

[0031] Where i t f represents the input gate value. t Represents the forgetting gate, o t Indicates the output gate value, C t-1 C represents the state of the memory unit at the previous moment. t C represents the current state of the memory unit. t+1 Indicates the state of the memory unit at the next moment, x t h represents the input of the current cell. t h represents the output of the cell at the current time. t-1 W represents the output of the cell at the previous time step. i W f W o W g These are the input gate, forget gate, output gate, and memory unit weights, respectively, b. i b f b o b g These are the input gate, forget gate, output gate, and memory unit bias term, respectively, and tanh(·) is the hyperbolic tangent function.

[0032] Furthermore, the self-encoder AE mentioned in step 2 is specifically as follows:

[0033] The autoencoder (AE) uses a compression algorithm to first compress the input data into a lower-dimensional feature, and then uses this lower-dimensional feature to reconstruct the input data. The reconstructed data is the output of the autoencoder.

[0034] An autoencoder (AE) is an unsupervised neural network model consisting of two parts: an encoder, defined by the function h = σ.e (W1x+b1) represents the other, which is the decoder, represented by the function y = σ. d (W2h+b2) represents, where x is the input, y is the output, W1 and b1 are the encoding weights and biases, W2 and b2 are the decoding weights and biases, and σ e (·), σ d (·) is the node activation function of the encoding and decoding networks, i.e., the sigmoid function, and h is the hidden state.

[0035] Furthermore, step 3 involves using the preprocessed training dataset as input to the LSTM-AE anomaly detection model for reconstruction training. The parameters are then optimized using the backpropagation algorithm until the evaluation metric is optimal. Finally, the LSTM-AE anomaly detection model and its parameters are saved, as detailed below:

[0036] Step 3.1: Use the preprocessed data tensor as the input and output of the model at each time step, and use the backpropagation algorithm to train the model based on the data reconstruction.

[0037] Step 3.2: Select the mean absolute error in L1 loss as the model loss function, and use the Adam optimizer. Given the learning rate and number of training iterations, optimize the parameters and update the network weights using the backpropagation algorithm until the evaluation metric is optimal.

[0038] Step 3.3: Save the LSTM-AE anomaly detection model and parameters.

[0039] Furthermore, in step 4, the periodic feature signal of the mobile robot under normal operating conditions is reconstructed using the trained LSTM-AE anomaly detection model, the reconstruction error is calculated, and the GCI of the periodic feature signal is calculated based on the reconstruction error, as detailed below:

[0040] Step 4.1: Describe the state of the industrial mobile robot using periodic time series data X. Utilize the trained LSTM-AE anomaly detection model to obtain a reconstructed state estimate of the industrial mobile robot. Calculate the model reconstruction error using the absolute error formula:

[0041]

[0042] Where X represents the input sequence. This represents the reconstructed output sequence, where E represents the absolute error and L represents the sample length.

[0043] Step 4.2: Model the reconstruction error for all time points using a Gaussian distribution, solve for the expected value μ and standard deviation σ of the error vector, and obtain the Gaussian distribution of the reconstruction error.

[0044] Step 4.3: Calculate 95% of the GCI under normal operating conditions using the following formula:

[0045] [μ-2σ,μ+2σ]

[0046] Furthermore, in step 5, the periodic feature signal acquired in real time is fed into the trained LSTM-AE anomaly detection model. By comparing the reconstruction error output by the LSTM-AE anomaly detection model with the GCI, it is determined whether the periodic feature signal is abnormal, as detailed below:

[0047] Step 5.1: Input the real-time extracted periodic feature signal into the pre-trained LSTM-AE anomaly detection model to obtain the reconstructed sequence output;

[0048] Step 5.2: Obtain the reconstruction error vector at this time point according to the reconstruction error calculation method;

[0049] Step 5.3: Compare the reconstruction error with the GCI. If the error is within the GCI, it is a normal data point; otherwise, it is an abnormal data point.

[0050] Compared with the prior art, the present invention has the following significant advantages: (1) It utilizes the long-term sequence-dependent memory capability of the LSTM deep model and the unsupervised learning and feature extraction capability of the AE model, replaces the AE neurons with LSTM neurons, introduces LSTM neurons into the autoencoder architecture, constructs the LSTM-AE anomaly detection model, and uses it to train and reconstruct periodic feature signals in the fault-free state of the industrial mobile robot system, so as to better extract the temporal features of the periodic feature signals; (2) When identifying different time points of the periodic feature signals, it uses the reconstruction error of the normal periodic feature signals to calculate the GCI, and uses the GCI to determine the periodic signal anomaly, which avoids the limitations of the traditional fixed threshold identification of anomalies and improves the accuracy of periodic feature signal anomaly detection in industrial mobile robots. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating the method for detecting abnormal periodic characteristic signals of an industrial mobile robot according to the present invention.

[0052] Figure 2 This is a schematic diagram of the LSTM unit in an embodiment of the present invention.

[0053] Figure 3 This is a structural framework diagram of the LSTM-AE anomaly detection model in an embodiment of the present invention.

[0054] Figure 4 This is a schematic diagram of the anomaly detection results of the periodic voltage time series based on a mobile robot handling system in an example of the present invention. Detailed Implementation

[0055] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0056] Combination Figure 1 The present invention discloses a method for detecting abnormal periodic characteristic signals of industrial mobile robots, comprising the following steps:

[0057] Step 1: Obtain periodic time series data of the industrial mobile robot and perform data preprocessing to obtain the training dataset, validation dataset, and test dataset for the model, as detailed below:

[0058] Step 1.1: Obtain the periodic time series data of the industrial mobile robot, as follows:

[0059] Step 1.1.1: Select periodic time series data of industrial mobile robots under suitable working conditions, and extract the required data using the equal time interval method;

[0060] Step 1.1.2: Based on the periodicity of the periodic characteristic signal, the collected time series is divided into samples of fixed length, with each sample serving as a time unit and L being the length of a sample.

[0061] Step 1.1.3: Set N to be the sample length of the extracted state periodic time series, then the periodic time series X = [X1, X2, X3, ..., X...]. N The format after dividing by sample length is as follows:

[0062]

[0063] Step 1.2: Perform data preprocessing to obtain the training dataset, validation dataset, and test dataset for the model, as detailed below:

[0064] The data for each time unit are standardized using the MinMaxScaler() function and then converted into tensor form to obtain the model's training dataset Tr_D (sample length N1), validation dataset Va_D (sample length N1), and test dataset Te_D (sample length N3), in the following format:

[0065]

[0066] Where tensor(·) represents a tensor. To represent a sample of data, Represents the training dataset. This represents the validation dataset. Let N represent the test dataset, and let i and j represent the starting points of the validation dataset and the test dataset, respectively. The datasets are mutually exclusive, and N1 + N2 + N3 = N.

[0067] Step 2: Utilizing the ability of LSTM deep models to extract time series characteristics, and based on the autoencoder (AE) reconstruction principle, construct an LSTM-AE anomaly detection model, as follows:

[0068] By using LSTM units as neurons in the encoding and decoding process, and utilizing the control gates and memory gates in the LSTM units, and based on the autoencoder (AE) reconstruction principle, an LSTM-AE anomaly detection model is constructed, and the characteristics of time periodic sequences are effectively learned.

[0069] The LSTM deep model described above is as follows:

[0070] LSTM deep models are a special type of RNN network that introduces the concept of gates to update state variables. It is controlled by three gates: the input gate, the forget gate, and the output gate. The input gate controls the network's input, the forget gate controls the memory units, giving LSTM long-term memory capabilities, and the output gate controls the network's output. Therefore, the calculation formula for an LSTM unit is as follows:

[0071] i t =σ(W i *[h t-1 ,x t ]+b i )

[0072] f t =σ(W f *[h t-1 ,x t ]+b f )

[0073] o t =σ(W o *[h t-1 ,x t ]+b o )

[0074] C t =tanh(W g *[h t-1 ,x t ]+b g )

[0075] C t+1 =σ(f t *C t-1 +i t *C t )

[0076] h t =tanh(C t+1 )*o t

[0077] Where i t f represents the input gate value. t Represents the forgetting gate, o t Indicates the output gate value, C t-1 C represents the state of the memory unit at the previous moment. t C represents the current state of the memory unit. t+1 Indicates the state of the memory unit at the next moment, x t h represents the input of the current cell. t h represents the output of the cell at the current time. t-1 W represents the output of the cell at the previous time step. i W f W o W g These are the input gate, forget gate, output gate, and memory unit weights, respectively, b. i b f b o b g These are the input gate, forget gate, output gate, and memory unit bias term, respectively, and tanh(·) is the hyperbolic tangent function.

[0078] The self-encoder AE is described in detail below:

[0079] The autoencoder (AE) uses a compression algorithm to first compress the input data into a lower-dimensional feature, and then uses this lower-dimensional feature to reconstruct the input data. The reconstructed data is the output of the autoencoder.

[0080] An autoencoder (AE) is an unsupervised neural network model consisting of two parts: an encoder, defined by the function h = σ. e (W1x+b1) represents the other, which is the decoder, represented by the function y = σ. d (W2h+b2) represents, where x is the input, y is the output, W1 and b1 are the encoding weights and biases, W2 and b2 are the decoding weights and biases, and σ e (·), σ d (·) is the node activation function of the encoding and decoding networks, i.e., the sigmoid function, and h is the hidden state.

[0081] Step 3: Use the preprocessed training dataset as input to the LSTM-AE anomaly detection model for reconstruction training. Optimize the parameters using the backpropagation algorithm until the evaluation metric is close to optimal. Save the LSTM-AE anomaly detection model and its parameters, as detailed below:

[0082] Step 3.1: Use the preprocessed data tensor as the input and output of the model at each time step, and use the backpropagation algorithm to train the model based on the data reconstruction.

[0083] Step 3.2: Select the mean absolute error in L1 Loss as the model loss function, use the Adam optimizer, and optimize the parameters and update the network weights using the backpropagation algorithm under the given learning rate and number of training iterations until the evaluation index is close to the optimal.

[0084] Step 3.3: Save the LSTM-AE anomaly detection model and parameters.

[0085] Step 4: Using the trained LSTM-AE anomaly detection model, reconstruct the periodic feature signal of the mobile robot under normal operating conditions, calculate the reconstruction error, and calculate the GCI of the periodic feature signal based on the reconstruction error, as follows:

[0086] Step 4.1: Describe the state of the industrial mobile robot using periodic time series data X, and obtain the state estimate of the industrial mobile robot using the trained LSTM-AE anomaly detection model. The model reconstruction error is calculated using the absolute error formula as follows:

[0087]

[0088] Where X represents the input sequence. denoted as the reconstructed output sequence, E represents the absolute error, and L represents the sample length.

[0089] Step 4.2: Model the reconstruction error for all time points using a Gaussian distribution, solve for the expected value μ and standard deviation σ of the error vector, and obtain the Gaussian distribution of the reconstruction error.

[0090] Step 4.3: Calculate 95% of the GCI under normal operating conditions according to the following formula: [μ-2σ, μ+2σ].

[0091] Step 5: Input the real-time acquired periodic feature signal into the trained LSTM-AE anomaly detection model. By comparing the reconstruction error output by the LSTM-AE anomaly detection model with the GCI, determine whether the periodic feature signal is abnormal, as follows:

[0092] Step 5.1: Input the real-time extracted periodic feature signal into the pre-trained LSTM-AE anomaly detection model to obtain the reconstructed sequence output;

[0093] Step 5.2: Obtain the reconstruction error vector at this time point according to the reconstruction error calculation method;

[0094] Step 5.3: Compare the reconstruction error with the GCI. If the error is within the GCI, it is a normal data point; otherwise, it is an abnormal data point.

[0095] Example 1

[0096] This embodiment is based on the LSTM-AE anomaly detection model for periodic voltage time series anomaly detection in an industrial robot handling system.

[0097] The LSTM-AE anomaly detection model constructed in this embodiment has the following parameter settings: the data dimension of the model's input and output layers is 1, the number of features of the LSTM neuron is 256, the network structure is 2 layers, and the number of iterations is set to 100. The optimizer is Adam, the learning rate is 0.008, the loss function is nn.L1Loss, and the remaining parameters are set to their default values.

[0098] Model training was implemented using the PyTorch package from JupyterLab.

[0099] Combination Figure 1 The specific technical solution of this implementation case includes the following steps:

[0100] Step 1: Collect periodic time series data of the industrial mobile robot and perform data preprocessing to obtain the training dataset, validation dataset, and test dataset for the model, as detailed below:

[0101] Step 1.1: Select periodic voltage time series data of a normally operating industrial mobile robot. L is a sample length, and L is related to the voltage period. N is the sample length of the extracted periodic voltage time series.

[0102] The periodic voltage time series is then:

[0103] V T =[v T 1, v T 2, v T 3, ..., v T N ]

[0104] The format after partitioning according to sample length L is as follows:

[0105]

[0106] Step 1.2: The periodic voltage time series is standardized using the MinMaxScaler() function, and divided into training datasets (sample length N1), validation datasets (sample length N2), and test datasets (sample length N3) according to a certain ratio. These datasets are then converted into tensors, denoted as follows: The format is as follows:

[0107]

[0108] Where tensor(·) represents a tensor. To represent a sample of data, The training dataset representing periodic voltage time series. A validation dataset representing a periodic voltage time series. Let N1 represent the test dataset of the periodic voltage time series, and let i and j represent the starting points of the validation dataset and the test dataset of the periodic voltage time series, respectively. The datasets are mutually exclusive, and N1 + N2 + N3 = N.

[0109] Step 2: Utilize the time series modeling capabilities of the LSTM deep model and, based on the autoencoder (AE) reconstruction principle, construct an LSTM-AE anomaly detection model, as follows:

[0110] Combination Figure 2 LSTM deep models are a special type of RNN network that introduces the concept of gates to update state variables. It is controlled by three gates: the input gate, the forget gate, and the output gate. The input gate controls the network's input, the forget gate controls the memory units, giving LSTM long-term memory capabilities, and the output gate controls the network's output. Therefore, the calculation formula for an LSTM unit is as follows:

[0111] i t =σ(W i *[h t-1 ,x t ]+b i )

[0112] f t =σ(W f *[h t-1 ,x t ]+b f )

[0113] o t =σ(W o *[h t-1 ,x t ]+b o )

[0114] C t =tanh(W g *[h t-1 ,x t ]+b g )

[0115] C t+1 =σ(f t *C t-1 +i t *C t )

[0116] h t=tanh(C t+1 )*o t

[0117] Where i t f represents the input gate value. t Represents the forgetting gate, o t Indicates the output gate value, C t-1 C represents the state of the memory unit at the previous moment. t C represents the current state of the memory unit. t+1 Indicates the state of the memory unit at the next moment, x t h represents the input of the current cell. t h represents the output of the cell at the current time. t-1 W represents the output of the cell at the previous time step. i W f W o W g These are the input gate, forget gate, output gate, and memory unit weights, respectively, b. i b f b o b g These are the input gate, forget gate, output gate, and memory unit bias term, respectively, and tanh(·) is the hyperbolic tangent function.

[0118] AE is an unsupervised neural network model that consists of two parts: an encoder and an encoder. The encoder is defined by the function h = σ. e (W1x+b1) represents the other, which is the decoder derived from the function y = σ. d (W2h+b2) represents, where x is the input, y is the output, W1 and b1 are the encoding weights and biases, W2 and b2 are the decoding weights and biases, and σ e (·), σ d (·) is the node activation function of the encoding and decoding networks, i.e., the sigmoid function, and h is the hidden state. The AE process can be described as follows: first, the input data is compressed into a lower-dimensional feature, and then the input data is reproduced using this lower-dimensional feature. The reproduced data is the output of the autoencoder. In essence, it is a compression algorithm.

[0119] Leveraging the ability of LSTM deep models to handle long-term series data, and combining this with the advantages of AE in reconstructing time series data in unsupervised learning, a model is constructed as follows: Figure 3 The LSTM-AE anomaly detection model shown is trained and reconstructed using periodic voltage time series data of an industrial mobile robot under normal conditions.

[0120] Step 3: Use the preprocessed training dataset as input to the LSTM-AE anomaly detection model for reconstruction training. Optimize the parameters using the backpropagation algorithm until the evaluation metric is close to optimal. Save the LSTM-AE anomaly detection model and its parameters, as detailed below:

[0121] Step 3.1: Use the preprocessed data tensor as the input and output of the model at each time step, and use the backpropagation algorithm to train the model based on the data reconstruction.

[0122] Step 3.2: Select the mean absolute error in L1 loss as the model loss function, and use the Adam optimizer. Given the learning rate and number of training iterations, optimize the parameters and update the network weights using the backpropagation algorithm until the evaluation metric is close to optimal.

[0123] Step 3.3: Save the LSTM-AE anomaly detection model and parameters.

[0124] Step 4: Using the trained LSTM-AE anomaly detection model, reconstruct the periodic voltage time series of the mobile robot under normal operating conditions, calculate the reconstruction error, and calculate the GCI of the periodic voltage time series based on the reconstruction error, as follows:

[0125] Step 4.1: Describe the state of the industrial mobile robot using periodic voltage time series data X. Utilize the trained LSTM-AE anomaly detection model to obtain the reconstructed state of the industrial mobile robot. Calculate the model reconstruction error using the absolute error formula:

[0126]

[0127] Where V T This represents the input periodic voltage time series. E represents the periodic voltage time series of the reconstructed output. V L represents the absolute error of the periodic voltage time series. V This represents the length of a single periodic voltage time series sample.

[0128] Step 4.2: Model the reconstruction error for all time points using a Gaussian distribution, solve for the expected value μ and standard deviation σ of the error vector E, and obtain the Gaussian distribution of the reconstruction error.

[0129] Step 4.3: Calculate 95% GCI under normal operating conditions using the following formula:

[0130] [μ-2σ,μ+2σ]

[0131] Step 5: Input the real-time acquired periodic voltage time series into the trained LSTM-AE anomaly detection model. By comparing the reconstruction error output by the LSTM-AE anomaly detection model with the GCI, determine whether the periodic voltage time series has anomalies, as follows:

[0132] Step 5.1: Input the real-time extracted periodic voltage time series into the pre-trained LSTM-AE anomaly detection model to obtain the reconstructed sequence output;

[0133] Step 5.2: Obtain the reconstruction error vector of the sequence points according to the reconstruction error calculation method;

[0134] Step 5.3: Compare the reconstruction error with the GCI. If the error is within the GCI, it is a normal data point; otherwise, it is an abnormal data point.

[0135] This embodiment uses the periodic voltage time series of an AGV transport vehicle in an industrial enterprise for testing and compares it with other popular periodic feature signal anomaly detection models. To verify the anomaly detection effect of the periodic feature signal based on the LSTM-AE framework, this embodiment conducts extensive experimental verification: First, the effectiveness of the LSTM-AE anomaly detection model in extracting periodic feature signals is verified by comparing the periodic signal anomaly detection based on the LSTM-AE framework with AE and VAE models. The confusion matrix shows the results of different models, and further evaluation indicators are obtained, including accuracy, recall, and false alarm rate (FA). The confusion matrix is ​​shown in Table 1, and the comparison results are shown in Table 2. Second, the efficiency of GCI in identifying outliers is verified by comparing the GCI based on the LSTM-AE anomaly detection model with traditional threshold anomaly identification. Considering the limitations and time consumption of traditional thresholds, the time consumption ratio T_Rate and threshold applicability G_Rate of the optimal threshold are selected as evaluation indicators. The comparison results are shown in Table 3.

[0136] Table 1 Confusion Matrix

[0137]

[0138]

[0139]

[0140]

[0141]

[0142]

[0143] Where TP is the number of samples that are actually abnormal when judged as abnormal, FN is the number of samples that are actually abnormal when judged as normal, FP is the number of samples that are actually normal when judged as abnormal, and TN is the number of samples that are actually normal when judged as normal.

[0144] Using 5 as the base, T is the total number of attempts required to find the optimal threshold. L E L is the length of the reconstruction error interval of the training dataset. G It refers to the GCI length. Based on the definition of the index, in a set of experimental results, the higher the accuracy and recall, the better; the lower the false alarm rate, the better; the lower the time-to-response ratio, the better; and the greater the applicability, the better.

[0145] Table 2 Anomaly detection results of different models

[0146] Accuracy Recall FA AE 56.5％ 45.3％ 12.5％ VAE 68.4％ 93.2％ 100％ LSTM-AE 96.0％ 95.7％ 5.3％

[0147] Table 3. Anomaly detection results at different thresholds

[0148]

[0149] Table 2 shows that the periodic signal anomaly detection model based on LSTM-AE outperforms the AE and VAE models, achieving a recall rate of up to 95% while maintaining high accuracy, with negligible false alarms. Table 3 shows that, considering the theoretical support and time consumption for adjusting to the optimal threshold, the GCI threshold range has advantages such as strong theoretical basis, short processing time, and wide applicability, even with a slight fluctuation in accuracy, which is negligible. Therefore, the LSTM-AE-based method for detecting periodic feature signals in industrial mobile robots presents the best overall performance. Combined with the high fidelity of GCI, it addresses the shortcomings of anomaly detection in industrial mobile robot systems. Leveraging the advantages of deep learning in big data analysis, it mines operational and maintenance data from industrial systems, providing crucial support for anomaly detection and early warning in mobile robot systems.

[0150] The embodiments described above are merely illustrative of design ideas and features, intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The scope of protection of the present invention is not limited to the above embodiments. All equivalent changes or modifications made based on the principles and design concepts disclosed in the present invention are within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for detecting anomalies in periodic characteristic signals of an industrial mobile robot, characterized in that, Includes the following steps: Step 1: Obtain periodic time series data of industrial mobile robots and perform data preprocessing to obtain training dataset, validation dataset and test dataset for the model; Step 2: Utilize the ability of the LSTM deep model to extract time series characteristics, and construct an LSTM-AE anomaly detection model based on the autoencoder (AE) reconstruction principle. Step 3: Use the preprocessed training dataset as input to the LSTM-AE anomaly detection model for reconstruction training. Optimize the parameters using the backpropagation algorithm until the evaluation metric is optimal. Save the LSTM-AE anomaly detection model and its parameters. Step 4: Using the trained LSTM-AE anomaly detection model, reconstruct the periodic feature signal of the mobile robot under normal working conditions, calculate the reconstruction error, and calculate the GCI of the periodic feature signal based on the reconstruction error. Step 5: Input the real-time acquired periodic feature signal into the trained LSTM-AE anomaly detection model. By comparing the reconstruction error output by the LSTM-AE anomaly detection model with the GCI, determine whether the periodic feature signal is abnormal. Step 1 involves acquiring periodic time series data of the industrial mobile robot and performing data preprocessing to obtain the model's training dataset, validation dataset, and test dataset, as detailed below: Step 1.1: Select periodic voltage time series data of industrial mobile robots under suitable working conditions, and extract the required data using the equal time interval method; Step 1.2: Based on the periodicity of the periodic characteristic signal, the collected time series is divided into samples of fixed length, with each sample serving as a time unit and L being the sample length. Step 1.3: Set N to be the sample length of the extracted state periodic time series, then the periodic time series The format after dividing by sample length is as follows: Step 1.4: Standardize the data for each time unit using the MinMaxScaler() function, then convert it into tensor form to obtain the training dataset for the model. Validation dataset and test dataset Training dataset The sample length is Validate the dataset The sample length is Test dataset The sample length is The format is as follows: in Represents a tensor. To represent a sample of data, Represents the training dataset. This represents the validation dataset. This represents the test dataset. , This indicates the starting point of the validation dataset and the test dataset. The datasets are mutually exclusive, and ; Step 2 utilizes the ability of the LSTM deep model to extract time series characteristics and constructs an LSTM-AE anomaly detection model based on the autoencoder (AE) reconstruction principle, as detailed below: Using LSTM units as neurons in the encoding and decoding process, and utilizing the control gates and memory gates in the LSTM units, and based on the autoencoder AE reconstruction principle, an LSTM-AE anomaly detection model is constructed to learn the characteristics of time periodic sequences. The LSTM deep model described in step 2 is as follows: LSTM deep model is a type of RNN network that introduces the concept of gates to update state variables. It is controlled by three gates: the input gate, the forget gate, and the output gate. The input gate controls the network's input, the forget gate controls the memory units, giving LSTM long-term memory capabilities, and the output gate controls the network's output. The calculation formula for an LSTM unit is as follows: in Indicates the input gate value. Indicates the forgetting threshold. Indicates the output gate value. This indicates the state of the memory unit at the previous moment. This indicates the current state of the memory unit. Indicates the state of the memory unit at the next moment. Indicates the input of the current cell. This indicates the output of the cell at the current time. This represents the output of the cell at the previous time step. These are the input gate, forget gate, output gate, and memory unit weights, respectively. These are the input gate, forget gate, output gate, and memory unit bias term, respectively. It is a hyperbolic tangent function; The self-encoder AE mentioned in step 2 is as follows: The autoencoder (AE) uses a compression algorithm to first compress the input data into a lower-dimensional feature, and then uses this lower-dimensional feature to reconstruct the input data. The reconstructed data is the output of the autoencoder. An autoencoder (AE) is an unsupervised neural network model consisting of two parts: an encoder, which is composed of a function... The other is the decoder, which is a function. It means that, among them It is input, It is the output. , These are the weights and biases of the encoding. , These are the decoding weights and biases. These are the node activation functions of the encoding and decoding networks, that is... function, It is in a hidden state.

2. The method for detecting abnormal periodic characteristic signals of industrial mobile robots according to claim 1, characterized in that, Step 3 involves using the preprocessed training dataset as input to the LSTM-AE anomaly detection model for reconstruction training. The parameters are then optimized using the backpropagation algorithm until the evaluation metric is optimal. Finally, the LSTM-AE anomaly detection model and its parameters are saved. The details are as follows: Step 3.1: Use the preprocessed data tensor as the input and output of the model at each time step, and use the backpropagation algorithm to train the model based on the data reconstruction. Step 3.2: Select the mean absolute error in L1 loss as the model loss function, and use the Adam optimizer. Given the learning rate and number of training iterations, optimize the parameters and update the network weights using the backpropagation algorithm until the evaluation metric is optimal. Step 3.3: Save the LSTM-AE anomaly detection model and parameters.

3. The method for detecting abnormal periodic characteristic signals of industrial mobile robots according to claim 1, characterized in that, Step 4 describes the use of the trained LSTM-AE anomaly detection model to reconstruct the periodic feature signal of the mobile robot under normal operating conditions, calculate the reconstruction error, and then calculate the GCI of the periodic feature signal based on the reconstruction error, as detailed below: Step 4.1: Use periodic time series data The state of the industrial mobile robot is described, and the reconstructed state estimate of the industrial mobile robot is obtained using the trained LSTM-AE anomaly detection model. The model reconstruction error is calculated according to the absolute error formula: in Represents the input sequence. This indicates the reconstructed output sequence. This represents the absolute error, and L represents the sample length. Step 4.2: Model the reconstruction error at all time points using a Gaussian distribution and solve for the expectation of the error vector. and standard deviation The Gaussian distribution of the reconstruction error was obtained. ; Step 4.3: Calculate 95% of the GCI under normal operating conditions using the following formula: 。 4. The method for detecting abnormal periodic characteristic signals of industrial mobile robots according to claim 1, characterized in that, Step 5 involves feeding the real-time acquired periodic feature signals into the trained LSTM-AE anomaly detection model. By comparing the reconstruction error output by the LSTM-AE anomaly detection model with the GCI, it is determined whether the periodic feature signals exhibit anomalies. The details are as follows: Step 5.1: Input the real-time extracted periodic feature signal into the pre-trained LSTM-AE anomaly detection model to obtain the reconstructed sequence output; Step 5.2: Obtain the reconstruction error vector at this time point according to the reconstruction error calculation method; Step 5.3: Compare the reconstruction error with the GCI. If the error is within the GCI, it is a normal data point; otherwise, it is an abnormal data point.

Images