A Dynamic Pressure Recognition Method for Flexible Sensor Arrays Based on the Fusion of Pressure Change Trend Maps and Spatiotemporal Feature Networks

Abstract

This invention discloses a dynamic pressure recognition method for flexible sensor arrays based on the fusion of pressure change trend maps and spatiotemporal feature networks, belonging to the field of flexible sensor array signal processing and intelligent recognition technology. The method first uses a flexible pressure sensor array to collect continuous pressure data, constructs a pressure matrix sequence, and generates a corresponding pressure heatmap sequence. Then, it obtains a pressure change matrix through differential operations on pressure matrices at adjacent time points, extracts effective pressure change regions by combining threshold judgments, and constructs a pressure change trend map representing the direction, magnitude, and region of pressure change. Further, it fuses the pressure heatmap and the pressure change trend map to construct spatiotemporal feature input data. Finally, it inputs the fused continuous time-series samples into a spatiotemporal feature network for feature extraction and classification, realizing the recognition of dynamic pressure states such as pressure increase, pressure maintenance, and pressure decrease. This invention can effectively characterize the dynamic pressure change process, improve the spatiotemporal feature expression capability and dynamic pressure recognition accuracy, and enhance the system's anti-noise interference capability and robustness. It can be applied to fields such as smart wearables, medical rehabilitation monitoring, robot tactile perception, and human-computer interaction.

Description

Technical Field

[0001] This invention relates to the field of signal processing and intelligent recognition technology for flexible sensor arrays, and in particular to a dynamic pressure recognition method for flexible sensor arrays based on the fusion of pressure change trend maps and spatiotemporal feature networks. Background Technology

[0002] Flexible pressure sensor arrays, as a novel type of pressure sensing device, can achieve real-time measurement of pressure distribution on complex curved surfaces or localized areas of the human body, and have been widely used in fields such as smart wearables, rehabilitation medicine, and robotic haptics. However, existing flexible pressure sensor arrays still face several technical challenges in dynamic pressure recognition. Insufficient utilization of pressure change information: Existing methods mainly focus on the spatial distribution of single-frame pressure data or pressure heatmaps, lacking effective modeling of pressure change trends over continuous time intervals. Therefore, the accuracy in identifying pressure increases, decreases, or stable states is low. Limited noise resistance: Flexible sensors are affected by environmental interference, unstable contact, and inherent device noise in practical applications, leading to fluctuations in pressure measurements. Existing identification methods lack adaptive noise processing mechanisms, affecting the reliability of identification. Lack of intelligence in dynamic pressure identification methods: Traditional methods typically rely on threshold judgment or simple classifiers for pressure state identification, failing to fully utilize spatiotemporal features and making it difficult to accurately identify dynamic information such as pressure change rate, direction of change, and pressure trajectory. Weak integration of algorithms and hardware: Existing technologies often ignore the characteristics of flexible sensor arrays in algorithm design, failing to fully utilize the spatiotemporal features of multi-frame continuous pressure data, resulting in unsatisfactory pressure pattern recognition performance and limiting its application in smart wearables, rehabilitation training, and robotic tactile sensing.

[0003] Therefore, a dynamic pressure identification method is needed that can fully utilize continuous pressure data, effectively encode pressure change trends, perform intelligent analysis by combining spatiotemporal features, and possess noise resistance capabilities, so as to improve the accuracy and reliability of dynamic pressure identification of flexible sensor arrays in practical applications. Summary of the Invention

[0004] To address the problems of insufficient pressure change process representation capability, inadequate utilization of temporal information, and low dynamic pressure recognition accuracy in existing flexible sensor array dynamic pressure recognition technologies, this invention provides a flexible sensor array dynamic pressure recognition method based on the fusion of pressure change trend map and spatiotemporal feature network.

[0005] This invention constructs a pressure change trend map to characterize the direction and degree of pressure change, and combines it with a spatiotemporal feature network to jointly model continuous pressure data, thereby achieving high-precision identification of dynamic pressure states and improving the dynamic sensing capability and identification reliability of flexible pressure sensing arrays in complex environments.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A dynamic pressure identification method for flexible sensor arrays based on the fusion of pressure change trend maps and spatiotemporal feature networks includes the following steps: Step S1: Collect continuous pressure data of the target area using a flexible pressure sensor array, and obtain a continuous pressure matrix sequence according to a preset sampling period.

[0007] Step S2: Generate a corresponding pressure heat map sequence based on the continuous pressure matrix sequence, which is used to characterize the spatial distribution of pressure in the target area at different times.

[0008] Step S3: Perform a difference operation on the pressure matrix at adjacent time points to obtain the pressure change matrix; construct a pressure change trend map based on the pressure change matrix to characterize the direction, magnitude, and region of pressure change. Before constructing the pressure change trend map, perform a threshold determination process on the pressure change matrix: set a change threshold δ; when the absolute value of the corresponding element in the pressure change matrix is ​​greater than δ, it is determined to be an effective pressure change region; otherwise, it is determined to be a pressure stable region.

[0009] Step S4: Integrate the pressure heat map with the pressure change trend map to construct multidimensional spatiotemporal feature data.

[0010] Step S5: Channel fusion of the pressure heat map sequence and the pressure trend map sequence to construct spatiotemporal feature input data.

[0011] Step S6: Input the spatiotemporal feature input data into a three-dimensional convolutional neural network for feature extraction.

[0012] Step S7: Complete the dynamic pressure state identification based on the output results of the three-dimensional convolutional neural network. The dynamic pressure state includes at least: pressure increase state, pressure maintenance state, and pressure decrease state.

[0013] Furthermore, the pressure change matrix in step S3 is obtained using the following formula: D(t) = P(t) - P(t-1) where: P(t) represents the pressure matrix at the current moment; P(t-1) represents the pressure matrix at the previous moment; D(t) represents the pressure change matrix.

[0014] Furthermore, the pressure change trend map is generated by normalizing the pressure change matrix and applying pseudo-color mapping to enhance the visual representation of pressure change regions.

[0015] Furthermore, the pressure change trend diagram includes: a pressure increase region; a pressure decrease region; and a pressure stability region.

[0016] Furthermore, the fusion methods in step S4 include: feature channel fusion; feature-level fusion; and decision-level fusion. The preferred approach is to use a feature channel fusion method, which combines the pressure heat map and the pressure change trend map as network inputs.

[0017] Furthermore, the spatiotemporal feature network employs one or more of the following: a three-dimensional convolutional neural network, a combination of a convolutional neural network and a long short-term memory network, or a combination of a convolutional neural network and a Transformer.

[0018] Furthermore, the dynamic pressure states include: pressure increasing state; pressure maintaining state; and pressure decreasing state.

[0019] Furthermore, the dynamic pressure state can be further divided into: rapid increase state; slow increase state; stable state; slow decrease state; and rapid decrease state.

[0020] This invention also provides a flexible sensor array dynamic pressure recognition system based on the fusion of pressure change trend map and spatiotemporal feature network, comprising: a data acquisition module; a heat map generation module; a trend map construction module; a feature fusion module; a spatiotemporal feature extraction module; a dynamic pressure recognition module; and a result output module. The aforementioned modules work together to execute the dynamic pressure identification method. Specifically, the data acquisition module executes step S1, the heatmap generation module executes step S2, the trend map construction module executes the trend map construction portion of steps S3 and S4, the feature fusion module executes step S5, the spatiotemporal feature extraction module executes step S6, the dynamic pressure identification module executes step S7, and the result output module outputs the pressure change category identification result. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the overall process of the dynamic pressure identification method of the present invention. Figure 2 A flowchart is constructed to illustrate the pressure change trend of this invention; Figure 3 This is a schematic diagram illustrating the fusion of pressure thermogram and trend graph of the present invention; Figure 4 This is a flowchart of the spatiotemporal sample construction process of the present invention; Figure 5 This is a schematic diagram of the spatiotemporal network backbone of the present invention. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings.

[0023] Example 1: The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Those skilled in the art should understand that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0024] The following detailed description illustrates the specific implementation method: This embodiment provides a dynamic pressure recognition method for flexible sensor arrays based on the fusion of pressure change trend maps and spatiotemporal feature networks. The overall process is as follows: Figure 1 As shown, the process includes steps such as pressure data acquisition, pressure change trend map construction, spatiotemporal feature fusion, and dynamic pressure state identification.

[0025] In this embodiment, the flexible sensor array refers to a thin-film pressure sensor made of flexible material, arranged in an array, specifically a 4×4 layout, with a single-point pressure measurement range of 0 to 5 kg and an operating temperature of -50℃ to +50℃.

[0026] Example 1 S1. Continuous pressure data acquisition A flexible pressure sensor array serves as the pressure sensing unit.

[0027] In this embodiment, the flexible pressure sensing array consists of 4×4 pressure sensing units, containing a total of 16 pressure detection nodes, with each sensing unit arranged in a matrix.

[0028] The system collects pressure data at a fixed sampling frequency, obtaining a pressure matrix at each sampling time: P(t) = [pij(t)]4×4. Furthermore, for an m×n flexible pressure sensing array, the pressure matrix at time t is expressed as: P(t) = [pij(t)]m×n Where: pij(t) represents the pressure value collected by the sensing unit in the i-th row and j-th column at time t; i=1,2,...,m; j=1,2,...,n.

[0029] The pressure matrix corresponding to a flexible pressure array of any size can be obtained using the above method.

[0030] Where: P(t) represents the pressure matrix at time t; pij(t) represents the pressure value of the sensing unit in the i-th row and j-th column at time t; t=1,2,3,…,T.

[0031] The pressure matrix sequence {P(1),P(2),...,P(T)} is continuously acquired and used to characterize the dynamic process of pressure change in the target area over time.

[0032] S2. Generation of Pressure Heatmap The pressure matrix collected at each time step is normalized: Pnorm(t)=(P(t)-Pmin) / (Pmax-Pmin) Where: Pmax is the maximum stress value of the current dataset; Pmin is the minimum stress value of the current dataset.

[0033] The normalized pressure matrix is ​​mapped to a pressure heatmap: the H(t) heatmap is generated using a pseudo-color mapping method.

[0034] In this design, high-pressure areas are represented by warm colors, while low-pressure areas are represented by cool colors. This results in a continuous pressure heatmap sequence: {H(1),H(2),...,H(T)}, used to describe the spatial distribution of pressure.

[0035] S3. Construction of Pressure Change Trend Chart To characterize the dynamic changes in pressure, this invention further constructs a pressure change trend graph.

[0036] First, calculate the pressure matrix difference between adjacent time points: D(t) = P(t) - P(t-1), where D(t) represents the pressure change matrix.

[0037] When D(t) > 0, it indicates that the pressure has increased; When D(t) < 0, it indicates that the pressure decreases; When D(t)=0, it indicates that the pressure remains stable.

[0038] The pressure variation matrix is ​​then normalized: Dnorm(t) = (D(t) - Dmin) / (Dmax - Dmin) The normalized matrix is ​​then used to generate a pressure change trend graph: T(t) through pseudo-color mapping. The red area represents areas of increased pressure; the blue area represents areas of decreased pressure; and the green area represents areas of stable pressure.

[0039] A continuous sequence of pressure change trend graphs is obtained: {T(2),T(3),...,T(T)}, which reflects the direction and magnitude of pressure changes. The pressure change trend graphs are also used to characterize the rate and intensity of pressure changes, wherein: the rate of change is calculated based on the ratio of the pressure difference between adjacent time points to the sampling time interval; the intensity of change is quantified using the absolute value or normalized amplitude of the change matrix.

[0040] S4. Spatiotemporal Feature Fusion The pressure heat map is integrated with the pressure change trend map.

[0041] Specifically, the heatmap H(t) and trend chart T(t) corresponding to the same moment are stacked according to the channel dimension: X(t) = Concat(H(t), T(t)), where X(t) represents the fused multi-channel feature data. To extract time-series features during continuous pressure changes, a sliding window approach is used to construct time-series samples: Xin = {X(t-N+1), ..., X(t)} Where: N represents the length of the time window.

[0042] In this embodiment, N is preferably set to 8 to 32 frames.

[0043] S5. Spatiotemporal Feature Network Construction and Feature Extraction The constructed time-series samples are input into the spatiotemporal feature network for training and recognition.

[0044] This embodiment uses a three-dimensional convolutional neural network as the spatiotemporal feature extraction network.

[0045] The network input dimensions are: N×H×W×C Where: N represents the length of the time dimension; H represents the image height; W represents the image width; and C represents the number of input channels.

[0046] The spatiotemporal feature network includes: an input layer; a 3D convolutional layer; a batch normalization layer; an activation layer; a pooling layer; and a fully connected layer.

[0047] After multiple convolution operations, a high-dimensional spatiotemporal feature vector is obtained: F This feature vector simultaneously includes: spatial distribution characteristics of pressure; trend characteristics of pressure changes; and temporal evolution characteristics.

[0048] S6. Dynamic Pressure Status Recognition The feature vector F is input into the classifier for state recognition.

[0049] The classifier uses a Softmax output layer: y = Softmax(F) Output dynamic pressure status category.

[0050] In this embodiment, the dynamic pressure states include: pressure increase; pressure maintenance; and pressure decrease.

[0051] Furthermore, it can be expanded to include: rapidly increasing; slowly increasing; stable; slowly decreasing; rapidly decreasing.

[0052] The final output is a dynamic pressure state sequence: Y={y1,y2,...,yn}, which enables real-time identification of pressure change processes in the target area.

[0053] This invention identifies continuous pressure data acquired by a flexible sensing array. Compared to methods that rely solely on pressure heatmaps, this invention enhances the expressive power of dynamic pressure changes by introducing a pressure trend map. This allows for a more intuitive and accurate representation of different dynamic states, such as pressure increase, pressure decrease, and pressure stability. Furthermore, by constructing the pressure trend map, key information such as the direction, magnitude, and region of change in the original pressure data is explicitly extracted, overcoming the limitation of traditional pressure heatmaps, which only reflect static pressure distribution and fail to adequately describe dynamic changes.

[0054] Furthermore, this invention employs a spatiotemporal feature network to jointly model pressure heatmaps and pressure change trend maps. By simultaneously extracting the spatial distribution and temporal evolution features of pressure data, it achieves deep feature learning of continuous pressure change processes. Compared to traditional methods based on single-frame image analysis or simple time-series statistics, this invention can more accurately capture the spatiotemporal correlations in the pressure change process, improve the distinguishability between different dynamic pressure states, and thus effectively enhance the accuracy of dynamic pressure identification.

[0055] Furthermore, since the pressure change trend map can highlight the pressure change area and suppress invalid background information, it can still maintain good feature representation ability under complex conditions such as sensor noise, environmental interference, and pressure fluctuations. Combined with the automatic feature extraction capability of the spatiotemporal feature network, this invention can effectively reduce the impact of noise on the recognition results and improve the system's anti-interference ability and robustness.

[0056] Example 2 Building upon Example 1, to fully extract the spatial distribution and temporal evolution features during pressure changes, this example constructs a spatiotemporal feature network based on a three-dimensional convolutional neural network (3D-CNN). Alternatively, the spatiotemporal feature network can employ a combination of a convolutional neural network and a long short-term memory network, where CNN is used to extract spatial features and LSTM is used to model temporal dependencies; or a combination of CNN and Transformer can be used, utilizing a self-attention mechanism to capture long-term pressure change features.

[0057] I. Network Input Construction First, the pressure heatmap and pressure change trend map are merged using channel fusion: X(t) = Concat(H(t), T(t)). Where: H(t) represents the pressure heat map at time t; T(t) represents the pressure change trend map at time t; X(t) represents the fused feature map.

[0058] Time series samples are constructed using a sliding time window of length N: Xin={X(t-N+1),X(t-N+2),...,X(t)} In this embodiment: N=16. Therefore, the network input tensor is represented as: 16×4×4×2 Where: 16 represents the time dimension; 4×4 represents the size of the flexible pressure array; 2 represents the heat map channel and the trend map channel.

[0059] II. Spatiotemporal Feature Network Structure The spatiotemporal feature network includes: an input layer; a first 3D convolutional layer; a first pooling layer; a second 3D convolutional layer; a second pooling layer; a fully connected layer; and a softmax classification layer.

[0060] III. First 3D Convolutional Layer The parameters of the first 3D convolutional layer are as follows: number of convolutional kernels: 32; kernel size: 3×3×3; stride: 1; padding method: SamePadding; input features: 16×4×4×2; output features: 16×4×4×32; this layer is mainly used to extract primary spatiotemporal features of pressure changes.

[0061] IV. The first pooling layer uses max pooling: MaxPool3D; pooling window: 2×2×2; used to reduce feature dimensions and improve network robustness.

[0062] Fifth, the second three-dimensional convolutional layer has 64 convolutional kernels, a kernel size of 3×3×3, a stride of 1, and an output feature of 8×2×2×64. This layer further extracts high-level pressure change features.

[0063] VI. The second pooling layer uses max pooling: MaxPool3D; pooling window: 2×2×2; to obtain compressed high-dimensional features.

[0064] 7. Fully connected layer: Unfolds the output of the convolutional network: Flatten; forms a one-dimensional feature vector: F; then inputs it into the fully connected layer: FC(128); outputs a 128-dimensional feature vector. This feature vector simultaneously contains: pressure spatial distribution features; pressure change trend features; and pressure temporal evolution features.

[0065] 8. Classification layer, using the Softmax classifier: y = Softmax(WF + b) Where: W is the weight matrix; b is the bias term; F is the spatiotemporal feature vector.

[0066] Output pressure state category: y∈{C1,C2,C3}, where: C1 represents pressure increase; C2 represents pressure maintenance; C3 represents pressure decrease.

[0067] IX. Network Training Process: Supervised learning was used to train the network. Training data was acquired using a flexible sensor array. Each sample group corresponds to a pressure state label. The labels are defined as follows: Category 1: Pressure increases; Category 2: Pressure remains constant; Category 3: Pressure decreases. The loss function used is the cross-entropy loss function: Loss = -∑yi·log(pi), where yi represents the true label and pi represents the predicted probability.

[0068] The optimizer used is the Adam optimizer.

[0069] The learning rate is preferably set to 0.0001.

[0070] The preferred batch size is 32.

[0071] The EarlyStopping strategy is used during training to prevent network overfitting.

[0072] Example 3 To fully characterize the dynamic changes in pressure during the detection process of a flexible pressure sensing array, this embodiment proposes a method for constructing a pressure change trend map based on the differential operation of a continuous pressure matrix.

[0073] This method extracts the change information between pressure data at adjacent time points to construct a pressure change trend map that reflects the direction, magnitude, and area of ​​pressure change, providing effective feature support for subsequent dynamic pressure identification.

[0074] I. Obtaining the Continuous Pressure Matrix Continuous data acquisition is performed using a flexible pressure sensor array at a preset sampling period.

[0075] Let the pressure matrix collected at time t be: in: Represents the pressure matrix at time t; This indicates the pressure value of the corresponding sensing unit; m represents the number of array rows; n represents the number of columns in the array.

[0076] This embodiment uses a 4×4 flexible pressure sensor array, therefore: Continuous data acquisition forms a pressure matrix sequence: II. Calculation of Pressure Change Matrix To extract information about pressure changes, differential calculations are performed on the pressure matrices at adjacent time points.

[0077] The pressure change matrix is ​​represented as follows: in: This represents the pressure change matrix at time t; This represents the pressure matrix at the current moment; This represents the pressure matrix at the previous moment.

[0078] Differential calculations can be used to obtain the pressure changes of each sensing unit between adjacent time points.

[0079] when: This indicates an increase in pressure in the corresponding area; when: This indicates that the pressure in the corresponding area has decreased; when: This indicates that the pressure in the corresponding area remains stable.

[0080] III. Extraction of Pressure Variation Areas Considering the noise interference and sampling fluctuations that occur in flexible pressure sensors during practical applications, this embodiment employs a threshold determination method to extract the effective change region in order to improve the accuracy of pressure change detection. A change threshold δ is set.

[0081] When the following conditions are met: When the pressure changes in the corresponding sensing unit, it is determined that an effective pressure change has occurred.

[0082] When the following conditions are met: When the time is right, the corresponding sensing unit is determined to be in a stable state.

[0083] This yields a matrix of pressure variation regions.

[0084] This step can effectively reduce the impact of environmental noise and sensor drift on the recognition results.

[0085] IV. Extraction of Pressure Change Direction The direction of pressure change is determined by the positive and negative relationships of the elements in the pressure change matrix.

[0086] Specifically, it includes: Increased pressure area when: At that time, the corresponding area is defined as the pressure enhancement area.

[0087] Pressure Reduction Area when: At that time, the corresponding area is defined as the pressure reduction area.

[0088] Pressure stability region when: When this occurs, the corresponding region is defined as the pressure stable region.

[0089] This provides information about the direction of pressure change.

[0090] V. Generation of Pressure Change Trend Chart The pressure change matrix is ​​then normalized. in: This represents the maximum value of the pressure change matrix; This represents the minimum value of the pressure change matrix.

[0091] A pressure change trend graph was then generated based on the normalization results.

[0092] Pressure change trend graphs are used to characterize: Direction of pressure change; Pressure change range; Areas of pressure variation; Pressure change process.

[0093] In this embodiment, a pseudo-color mapping method is used for visualization.

[0094] in: The red area indicates increased pressure. The blue area indicates reduced pressure; The green area indicates that the pressure remains stable.

[0095] The final pressure change trend chart for the corresponding time point is obtained: A series of trend charts are formed over consecutive time periods: VI. Output of Pressure Change Trend Chart The generated pressure change trend map is used as a dynamic pressure change feature input into the subsequent spatiotemporal feature network.

[0096] This data, together with the corresponding pressure heatmap at that time, constitutes the fused input data. in: Represents a pressure heat map; This graph represents the trend of pressure changes.

[0097] The fused input data is used for subsequent dynamic pressure state identification.

[0098] Example 4 Example of a method for fusing pressure thermograms and pressure change trend graphs Based on Examples 1 to 3, in order to make full use of pressure spatial distribution information and pressure dynamic change information, this example proposes a method for fusing pressure heat map and pressure change trend map.

[0099] This method improves the accuracy of dynamic pressure state identification by constructing a multi-channel spatiotemporal feature expression model to achieve joint representation of static and dynamic pressure information.

[0100] I. Data Source Integration The merged data includes: (1) Pressure-thermal diagram; (2) Pressure change trend graph.

[0101] in: Pressure heatmaps are used to characterize the spatial distribution of pressure in a target area at a given moment. Pressure change trend graphs are used to characterize the direction, magnitude, and area of ​​pressure change.

[0102] For time t: The pressure-thermal diagram is represented as: H(t) The pressure change trend is represented by: T(t) Both have the same spatial dimensions.

[0103] In this embodiment: H(t)∈R4×4; T(t)∈R4×4 II. Data Normalization Processing To eliminate the influence of different dimensions, the pressure heat map and pressure change trend map were normalized respectively.

[0104] The normalized representation of the pressure-thermal diagram is: H'(t) The normalized representation of the pressure change trend graph is: T'(t) After normalization: all eigenvalues ​​are mapped to the interval [0,1].

[0105] This ensures that different features have the same data scale.

[0106] III. Feature Channel Fusion This embodiment preferably uses a feature channel fusion method.

[0107] Stack the pressure heat map and the pressure change trend map at the same moment along the channel direction: X(t)=Concat(H'(t),T'(t)) Wherein: Concat represents the channel splicing operation; X(t) represents the fused feature map.

[0108] The fused feature dimension is represented as: X(t)∈R4×4×2 The first channel represents the pressure heatmap feature, and the second channel represents the pressure change trend map feature. This structure enables the simultaneous expression of static and dynamic pressure features.

[0109] IV. Construction of Time Series Samples To extract temporal correlation features during pressure changes, a sliding window mechanism was used to construct continuous time-series samples.

[0110] Set the window length to N.

[0111] The network input consists of N consecutive frames of fused feature maps: S(t) = {X(t-N+1), X(t-N+2), ..., X(t)} In this embodiment, N is preferably set to 16.

[0112] Therefore, the input tensor is formed as: 16×4×4×2 Where: 16 represents the time dimension; 4×4 represents the spatial dimension; 2 represents the number of fusion channels.

[0113] V. Multi-layer fusion feature representation The fused features contain the following information: The first type of feature is the spatial distribution of pressure, which is used to reflect the stress state of the target area.

[0114] The second type of characteristic: the direction of pressure change; used to reflect the process of pressure increasing or decreasing.

[0115] The third type of characteristic: pressure change amplitude; used to reflect the intensity of pressure change.

[0116] The fourth type of characteristic: pressure time evolution characteristics; used to reflect the continuous change of pressure over time.

[0117] Therefore, a complete spatiotemporal feature representation model of pressure is constructed.

[0118] VI. Fusion Feature Input Spatiotemporal Feature Network The constructed fused features are input into the spatiotemporal feature network: F = Network(S) Where: Network represents the spatiotemporal feature network; F represents the output spatiotemporal feature vector.

[0119] By automatically learning the relationships between different features through the network, deep feature extraction of dynamic stress patterns can be achieved.

[0120] VII. Analysis of the Advantages of the Integration Mechanism Compared to methods that rely solely on pressure thermograms for identification, this embodiment offers the following advantages: (1) It can simultaneously retain information on the spatial distribution of pressure and information on pressure changes; (2) Improve the separability of characteristics between dynamic pressure states; (3) Enhance the network's ability to perceive pressure changes; (4) Reduce the impact of static background information on the recognition results; (5) Improve the ability to identify subtle pressure changes.

[0121] Experimental results show that by fusing the pressure heatmap and the pressure change trend map, the network can obtain richer spatiotemporal information about pressure.

[0122] Compared to methods that only use pressure heatmaps as input to the network, this embodiment demonstrates higher classification accuracy in all three dynamic pressure state recognition tasks: pressure increase, pressure maintenance, and pressure decrease. Especially under conditions of small pressure changes, complex changes, and noise interference, the fused features still maintain good discriminative ability and stability.

[0123] Therefore, the fusion method proposed in this embodiment can effectively improve the recognition performance and engineering application value of the flexible sensor array dynamic pressure recognition system.

[0124] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dynamic pressure identification method for flexible sensor arrays based on the fusion of pressure change trend maps and spatiotemporal feature networks, characterized in that, Includes the following steps S1. Use a flexible pressure sensor array to collect pressure data of the target area at continuous time intervals, and generate a continuous pressure matrix sequence according to a preset time interval. S2. Construct a corresponding pressure heat map sequence based on the continuous pressure matrix sequence; S3. Perform a difference operation on the pressure matrix at adjacent time points to obtain the pressure change trend matrix; S4. Construct a pressure change trend diagram based on the pressure change trend matrix. The pressure change trend diagram is used to characterize the direction, rate and intensity of pressure change. S5. Channel fusion of the pressure heat map sequence and the pressure change trend map sequence to construct spatiotemporal characteristic input data; S6. Input the spatiotemporal feature input data into a three-dimensional convolutional neural network for feature extraction; S7. Based on the output of the three-dimensional convolutional neural network, complete the dynamic pressure state identification. The dynamic pressure state includes at least: pressure increase state; pressure maintenance state; and pressure decrease state.

2. The dynamic pressure identification method for flexible sensor arrays based on the fusion of pressure change trend maps and spatiotemporal feature networks according to claim 1, characterized in that: The pressure matrix is: P(t) = [p ij [(t)], where: p ij (t) represents the pressure value of the sensing unit in the i-th row and j-th column at time t.

3. The dynamic pressure identification method for flexible sensor arrays based on the fusion of pressure change trend maps and spatiotemporal feature networks according to claim 1, characterized in that: The pressure change trend matrix is ​​obtained using the following formula: ； in: This is the pressure matrix at the current moment; This is the pressure matrix from the previous moment; This is a matrix showing the trend of pressure changes.

4. The dynamic pressure recognition method for flexible sensor arrays based on the fusion of pressure change trend map and spatiotemporal feature network according to claim 3, characterized in that: Before constructing the pressure change trend map, a threshold determination process is performed on the pressure change matrix; A change threshold δ is set. When the absolute value of the corresponding element in the pressure change matrix is ​​greater than the change threshold δ, it is determined to be an effective pressure change region; otherwise, it is determined to be a pressure stable region.

5. The dynamic pressure identification method for flexible sensor arrays based on the fusion of pressure change trend map and spatiotemporal feature network according to claim 1, characterized in that: The fusion method described in step S4 includes: Feature channel fusion; Feature-level fusion; Decision-level integration; The preferred approach is to use a feature channel fusion method, which stitches the pressure heat map and the pressure change trend map along the channel dimension as the network input.

6. The dynamic pressure identification method for flexible sensor arrays based on the fusion of pressure change trend maps and spatiotemporal feature networks according to any one of claims 1-5, characterized in that: The pressure heat map and pressure change trend map are normalized separately and then merged in the following manner: X(t) = Concat(H(t), T(t)); in: H(t) is a pressure-thermal diagram; T(t) is a graph showing the trend of pressure change; Concat is a channel concatenation operation; X(t) is the fused multi-channel feature map.

7. The dynamic pressure identification method for flexible sensor arrays based on the fusion of pressure change trend maps and spatiotemporal feature networks according to any one of claims 1-5, characterized in that: Time series samples are constructed using a sliding time window: S(t)={X(t-N+1),X(t-N+2),…,X(t)}; Where N represents the length of the time window; The time window length N is 8 to 32 frames.

8. The dynamic pressure identification method for flexible sensor arrays based on the fusion of pressure change trend map and spatiotemporal feature network according to claim 1, characterized in that: The spatiotemporal feature network is one or more of the following network structures: 3D convolutional neural network; A combined structure of convolutional neural networks and long short-term memory networks; A combination structure of convolutional neural networks and Transformers.

9. The dynamic pressure identification method for flexible sensor arrays based on the fusion of pressure change trend maps and spatiotemporal feature networks according to any one of claims 1 to 8, characterized in that: The dynamic pressure state includes: Increased pressure; Pressure maintenance status; Pressure decreases; Or further include: Rapidly increasing state; Slowly increasing state; steady state; Slowly decreasing state; The state is rapidly decreasing.

10. A flexible sensor array dynamic pressure identification system based on the fusion of pressure change trend map and spatiotemporal feature network, characterized in that, include: The data acquisition module is used to acquire continuous pressure matrix sequences; The heat map generation module is used to generate pressure heat map sequences; The trend chart building module is used to build trend charts of pressure changes; The feature fusion module is used to fuse the pressure heat map and the pressure change trend map; The spatiotemporal feature extraction module is used to extract the spatial and temporal features of pressure data; The dynamic pressure identification module is used to identify dynamic pressure status. The results output module is used to output the pressure change category; Each module works together to execute the method described in any one of claims 1 to 9.

Images